diff --git a/examples/ezmsg_configs.py b/examples/ezmsg_configs.py index 113295eb..4991acd8 100644 --- a/examples/ezmsg_configs.py +++ b/examples/ezmsg_configs.py @@ -82,8 +82,8 @@ async def listen(self, msg: int) -> None: class PassthroughCollection(ez.Collection): - INPUT = ez.InputStream(int) - OUTPUT = ez.OutputStream(int) + INPUT = ez.InputTopic(int) + OUTPUT = ez.OutputTopic(int) def network(self) -> ez.NetworkDefinition: return ((self.INPUT, self.OUTPUT),) @@ -136,7 +136,7 @@ def configure(self) -> None: class PubNoSubCollection(ez.Collection): - OUTPUT = ez.OutputStream(int) + OUTPUT = ez.OutputTopic(int) GENERATE = Generator() LOG = DebugLog() @@ -148,7 +148,7 @@ def network(self) -> ez.NetworkDefinition: class SubNoPubCollection(ez.Collection): - INPUT = ez.InputStream(int) + INPUT = ez.InputTopic(int) LISTEN = Listener() def network(self) -> ez.NetworkDefinition: @@ -175,7 +175,7 @@ class PubNoSubPassthroughCollection(ez.Collection): COLLECTION = PubNoSubCollection() PASSTHROUGH = PassthroughCollection() - OUTPUT = ez.OutputStream(int) + OUTPUT = ez.OutputTopic(int) def network(self) -> ez.NetworkDefinition: return ( @@ -188,7 +188,7 @@ class SubNoPubPassthroughCollection(ez.Collection): COLLECTION = SubNoPubCollection() PASSTHROUGH = PassthroughCollection() - INPUT = ez.InputStream(int) + INPUT = ez.InputTopic(int) def network(self) -> ez.NetworkDefinition: return ( diff --git a/examples/ezmsg_toy.py b/examples/ezmsg_toy.py index a5c5c772..154d401c 100644 --- a/examples/ezmsg_toy.py +++ b/examples/ezmsg_toy.py @@ -24,20 +24,33 @@ class LFOSettings(ez.Settings): update_rate: float = 2.0 # Hz, update rate +class LFOState(ez.State): + start_time: float + cur_settings: LFOSettings + + class LFO(ez.Unit): SETTINGS = LFOSettings + STATE = LFOState OUTPUT = ez.OutputStream(float) + INPUT_SETTINGS = ez.InputStream(LFOSettings) + async def initialize(self) -> None: - self.start_time = time.time() + self.STATE.cur_settings = self.SETTINGS + self.STATE.start_time = time.time() + @ez.subscriber(INPUT_SETTINGS) + async def on_settings(self, msg: LFOSettings) -> None: + self.STATE.cur_settings = msg + @ez.publisher(OUTPUT) async def generate(self) -> AsyncGenerator: while True: - t = time.time() - self.start_time - yield self.OUTPUT, math.sin(2.0 * math.pi * self.SETTINGS.freq * t) - await asyncio.sleep(1.0 / self.SETTINGS.update_rate) + t = time.time() - self.STATE.start_time + yield self.OUTPUT, math.sin(2.0 * math.pi * self.STATE.cur_settings.freq * t) + await asyncio.sleep(1.0 / self.STATE.cur_settings.update_rate) # MESSAGE GENERATOR @@ -45,17 +58,30 @@ class MessageGeneratorSettings(ez.Settings): message: str +class MessageGeneratorState(ez.State): + cur_settings: MessageGeneratorSettings + + class MessageGenerator(ez.Unit): SETTINGS = MessageGeneratorSettings + STATE = MessageGeneratorState OUTPUT = ez.OutputStream(str) + INPUT_SETTINGS = ez.InputStream(MessageGeneratorSettings) + + async def initialize(self) -> None: + self.STATE.cur_settings = self.SETTINGS + + @ez.subscriber(INPUT_SETTINGS) + async def on_settings(self, msg: MessageGeneratorSettings) -> None: + self.STATE.cur_settings = msg @ez.publisher(OUTPUT) async def spawn_message(self) -> AsyncGenerator: while True: await asyncio.sleep(1.0) - ez.logger.info(f"Spawning {self.SETTINGS.message}") - yield self.OUTPUT, self.SETTINGS.message + ez.logger.info(f"Spawning {self.STATE.cur_settings.message}") + yield self.OUTPUT, self.STATE.cur_settings.message @ez.publisher(OUTPUT) async def spawn_once(self) -> AsyncGenerator: @@ -123,8 +149,8 @@ class ModifierCollection(ez.Collection): """This collection will subscribe to messages and append the most recent LFO output""" - INPUT = ez.InputStream(str) - OUTPUT = ez.OutputStream(str) + INPUT = ez.InputTopic(str) + OUTPUT = ez.OutputTopic(str) SIN = LFO() # SIN2 = LFO() @@ -152,6 +178,8 @@ class TestSystemSettings(ez.Settings): class TestSystem(ez.Collection): SETTINGS = TestSystemSettings + OUTPUT_PING = ez.OutputTopic(str) + # Publishers PING = MessageGenerator() FOO = MessageGenerator() @@ -173,6 +201,7 @@ def configure(self) -> None: # Define Connections def network(self) -> ez.NetworkDefinition: return ( + (self.PING.OUTPUT, self.OUTPUT_PING), (self.PING.OUTPUT, self.PINGSUB1.INPUT), (self.PING.OUTPUT, self.MODIFIER_COLLECTION.INPUT), (self.MODIFIER_COLLECTION.OUTPUT, self.PINGSUB2.INPUT), @@ -193,7 +222,7 @@ def process_components(self): ez.run( SYSTEM=system, connections=[ - # Make PING.OUTPUT available on a topic ezmsg_attach.py - (system.PING.OUTPUT, "GLOBAL_PING_TOPIC"), + # Make a system output available on a topic ezmsg_attach.py + (system.OUTPUT_PING, "GLOBAL_PING_TOPIC"), ], ) diff --git a/examples/profiling_tui.py b/examples/profiling_tui.py new file mode 100644 index 00000000..3f49556f --- /dev/null +++ b/examples/profiling_tui.py @@ -0,0 +1,450 @@ +#!/usr/bin/env python3 +""" +Simple live profiling TUI for ezmsg GraphServer. + +Features: +- Periodic profiling snapshot view broken out by publisher/subscriber endpoints +- Live trace sample counts via GraphContext.subscribe_profiling_trace() +- Optional automatic trace enablement for discovered processes + +Usage: + .venv/bin/python examples/profiling_tui.py --host 127.0.0.1 --port 25978 +""" + +from __future__ import annotations + +import argparse +import asyncio +import contextlib +import time +from dataclasses import dataclass +from uuid import UUID + +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import ( + ProcessProfilingSnapshot, + ProfilingStreamControl, + ProfilingTraceControl, +) +from ezmsg.core.netprotocol import DEFAULT_HOST, GRAPHSERVER_PORT_DEFAULT + + +def _truncate(text: object, width: int) -> str: + text = str(text) + if width <= 3: + return text[:width] + if len(text) <= width: + return text + return text[: width - 3] + "..." + + +def _fmt_float(value: float, digits: int = 2) -> str: + return f"{value:.{digits}f}" + + +@dataclass +class PublisherView: + process_id: UUID + topic: str + endpoint_id: str + published_total: int + published_window: int + publish_rate_hz: float + publish_delta_ms_avg: float + inflight_current: int + inflight_peak: int + trace_samples_seen: int + trace_last_age_s: float | None + backpressure_wait_ms_window: float + + +@dataclass +class SubscriberView: + process_id: UUID + topic: str + endpoint_id: str + channel_kind: str + received_total: int + received_window: int + lease_time_ms_avg: float + user_span_ms_avg: float + attributable_backpressure_ms_window: float + attributable_backpressure_events_total: int + trace_samples_seen: int + trace_last_age_s: float | None + + +class ProfilingTUI: + def __init__( + self, + ctx: GraphContext, + *, + snapshot_interval: float, + trace_interval: float, + trace_max_samples: int, + auto_trace: bool, + trace_sample_mod: int, + max_rows: int, + ) -> None: + self.ctx = ctx + self.snapshot_interval = max(0.2, snapshot_interval) + self.trace_interval = max(0.01, trace_interval) + self.trace_max_samples = max(1, trace_max_samples) + self.auto_trace = auto_trace + self.trace_sample_mod = max(1, trace_sample_mod) + self.max_rows = max(5, max_rows) + + self.snapshots: dict[UUID, ProcessProfilingSnapshot] = {} + self.route_units: dict[UUID, str] = {} + self.trace_enabled_processes: set[UUID] = set() + self.trace_errors: dict[UUID, str] = {} + self.trace_samples_seen_by_endpoint: dict[str, int] = {} + self.trace_last_timestamp_by_endpoint: dict[str, float] = {} + self.last_snapshot_time: float | None = None + + self._snapshot_task: asyncio.Task[None] | None = None + self._trace_task: asyncio.Task[None] | None = None + + async def start(self) -> None: + await self._refresh_snapshot() + self._snapshot_task = asyncio.create_task(self._snapshot_loop()) + self._trace_task = asyncio.create_task(self._trace_loop()) + + async def close(self) -> None: + for task in (self._snapshot_task, self._trace_task): + if task is not None: + task.cancel() + with contextlib.suppress(asyncio.CancelledError): + await task + + if not self.auto_trace: + return + + for process_id, route_unit in self.route_units.items(): + if process_id not in self.trace_enabled_processes: + continue + with contextlib.suppress(Exception): + await self.ctx.process_set_profiling_trace( + route_unit, + ProfilingTraceControl(enabled=False), + timeout=0.5, + ) + + async def _snapshot_loop(self) -> None: + while True: + await self._refresh_snapshot() + await asyncio.sleep(self.snapshot_interval) + + async def _refresh_snapshot(self) -> None: + graph_snapshot = await self.ctx.snapshot() + route_units: dict[UUID, str] = {} + for process in graph_snapshot.processes.values(): + if process.units: + route_units[process.process_id] = process.units[0] + self.route_units = route_units + + if self.auto_trace: + for process_id, route_unit in route_units.items(): + if process_id in self.trace_enabled_processes: + continue + try: + response = await self.ctx.process_set_profiling_trace( + route_unit, + ProfilingTraceControl( + enabled=True, + sample_mod=self.trace_sample_mod, + ), + timeout=0.5, + ) + if response.ok: + self.trace_enabled_processes.add(process_id) + self.trace_errors.pop(process_id, None) + else: + self.trace_errors[process_id] = str( + response.error or "unknown error" + ) + except Exception as exc: + self.trace_errors[process_id] = str(exc) + + self.snapshots = await self.ctx.profiling_snapshot_all( + timeout_per_process=max(0.1, self.snapshot_interval * 0.8) + ) + self.last_snapshot_time = time.time() + + async def _trace_loop(self) -> None: + async for batch in self.ctx.subscribe_profiling_trace( + ProfilingStreamControl( + interval=self.trace_interval, + max_samples=self.trace_max_samples, + ) + ): + for process_batch in batch.batches.values(): + for sample in process_batch.samples: + endpoint_id = sample.endpoint_id + self.trace_samples_seen_by_endpoint[endpoint_id] = ( + self.trace_samples_seen_by_endpoint.get(endpoint_id, 0) + 1 + ) + self.trace_last_timestamp_by_endpoint[endpoint_id] = batch.timestamp + + def _trace_for_endpoint(self, endpoint_id: str) -> tuple[int, float | None]: + now = time.time() + count = self.trace_samples_seen_by_endpoint.get(endpoint_id, 0) + ts = self.trace_last_timestamp_by_endpoint.get(endpoint_id) + age = None if ts is None else max(0.0, now - ts) + return count, age + + def _publisher_rows(self) -> list[PublisherView]: + rows: list[PublisherView] = [] + for process_id, snapshot in self.snapshots.items(): + for pub in snapshot.publishers.values(): + trace_count, trace_age = self._trace_for_endpoint(pub.endpoint_id) + rows.append( + PublisherView( + process_id=process_id, + topic=pub.topic, + endpoint_id=pub.endpoint_id, + published_total=pub.messages_published_total, + published_window=pub.messages_published_window, + publish_rate_hz=pub.publish_rate_hz_window, + publish_delta_ms_avg=pub.publish_delta_ns_avg_window / 1_000_000.0, + inflight_current=pub.inflight_messages_current, + inflight_peak=pub.inflight_messages_peak_window, + trace_samples_seen=trace_count, + trace_last_age_s=trace_age, + backpressure_wait_ms_window=( + pub.backpressure_wait_ns_window / 1_000_000.0 + ), + ) + ) + rows.sort( + key=lambda row: ( + -row.publish_rate_hz, + -row.published_total, + row.process_id, + row.topic, + ) + ) + return rows + + def _subscriber_rows(self) -> list[SubscriberView]: + rows: list[SubscriberView] = [] + for process_id, snapshot in self.snapshots.items(): + for sub in snapshot.subscribers.values(): + trace_count, trace_age = self._trace_for_endpoint(sub.endpoint_id) + channel_kind = ( + sub.channel_kind_last.value + if hasattr(sub.channel_kind_last, "value") + else str(sub.channel_kind_last) + ) + rows.append( + SubscriberView( + process_id=process_id, + topic=sub.topic, + endpoint_id=sub.endpoint_id, + channel_kind=channel_kind, + received_total=sub.messages_received_total, + received_window=sub.messages_received_window, + lease_time_ms_avg=sub.lease_time_ns_avg_window / 1_000_000.0, + user_span_ms_avg=sub.user_span_ns_avg_window / 1_000_000.0, + attributable_backpressure_ms_window=( + sub.attributable_backpressure_ns_window / 1_000_000.0 + ), + attributable_backpressure_events_total=( + sub.attributable_backpressure_events_total + ), + trace_samples_seen=trace_count, + trace_last_age_s=trace_age, + ) + ) + rows.sort( + key=lambda row: ( + -row.lease_time_ms_avg, + -row.received_total, + row.process_id, + row.topic, + ) + ) + return rows + + def render(self) -> None: + print("\x1bc", end="") + print("ezmsg profiling tui") + print("Ctrl-C to quit") + print( + "snapshot interval=" + f"{self.snapshot_interval:.2f}s, trace interval={self.trace_interval:.2f}s, " + f"trace max_samples={self.trace_max_samples}, auto_trace={self.auto_trace}" + ) + if self.last_snapshot_time is not None: + print( + "last snapshot age: " + f"{_fmt_float(max(0.0, time.time() - self.last_snapshot_time), 2)}s" + ) + print( + f"processes discovered={len(self.route_units)} " + f"publishers={sum(len(s.publishers) for s in self.snapshots.values())} " + f"subscribers={sum(len(s.subscribers) for s in self.snapshots.values())}" + ) + + publisher_rows = self._publisher_rows() + subscriber_rows = self._subscriber_rows() + + print("\nPublishers") + pub_header = ( + f"{'Process':<20} {'Topic':<26} {'Endpoint':<24} " + f"{'Total':>8} {'Win':>6} {'RateHz':>8} {'DeltaMs':>8} " + f"{'InFl':>5} {'InPk':>5} {'BPmsW':>8} {'Trace':>7} {'TAge':>6}" + ) + print(pub_header) + print("-" * len(pub_header)) + if not publisher_rows: + print("") + else: + for row in publisher_rows[: self.max_rows]: + trace_age = ( + "-" if row.trace_last_age_s is None else _fmt_float(row.trace_last_age_s, 2) + ) + print( + f"{_truncate(row.process_id, 20):<20} " + f"{_truncate(row.topic, 26):<26} " + f"{_truncate(row.endpoint_id, 24):<24} " + f"{row.published_total:>8} " + f"{row.published_window:>6} " + f"{_fmt_float(row.publish_rate_hz, 2):>8} " + f"{_fmt_float(row.publish_delta_ms_avg, 2):>8} " + f"{row.inflight_current:>5} " + f"{row.inflight_peak:>5} " + f"{_fmt_float(row.backpressure_wait_ms_window, 2):>8} " + f"{row.trace_samples_seen:>7} " + f"{trace_age:>6}" + ) + + print("\nSubscribers") + sub_header = ( + f"{'Process':<20} {'Topic':<26} {'Endpoint':<24} {'Kind':<6} " + f"{'Total':>8} {'Win':>6} {'LeaseMs':>8} {'UserMs':>8} " + f"{'BPmsW':>8} {'BPev':>6} {'Trace':>7} {'TAge':>6}" + ) + print(sub_header) + print("-" * len(sub_header)) + if not subscriber_rows: + print("") + else: + for row in subscriber_rows[: self.max_rows]: + trace_age = ( + "-" if row.trace_last_age_s is None else _fmt_float(row.trace_last_age_s, 2) + ) + print( + f"{_truncate(row.process_id, 20):<20} " + f"{_truncate(row.topic, 26):<26} " + f"{_truncate(row.endpoint_id, 24):<24} " + f"{_truncate(row.channel_kind, 6):<6} " + f"{row.received_total:>8} " + f"{row.received_window:>6} " + f"{_fmt_float(row.lease_time_ms_avg, 2):>8} " + f"{_fmt_float(row.user_span_ms_avg, 2):>8} " + f"{_fmt_float(row.attributable_backpressure_ms_window, 2):>8} " + f"{row.attributable_backpressure_events_total:>6} " + f"{row.trace_samples_seen:>7} " + f"{trace_age:>6}" + ) + + if self.trace_errors: + print("\ntrace errors:") + for process_id, err in sorted(self.trace_errors.items(), key=lambda item: str(item[0])): + print(f" {_truncate(str(process_id), 30)}: {_truncate(err, 120)}") + + +def _parse_address(host: str, port: int) -> tuple[str, int]: + return (host, port) + + +async def _run_tui(args: argparse.Namespace) -> None: + async with GraphContext( + _parse_address(args.host, args.port), auto_start=args.auto_start + ) as ctx: + tui = ProfilingTUI( + ctx, + snapshot_interval=args.snapshot_interval, + trace_interval=args.trace_interval, + trace_max_samples=args.max_samples, + auto_trace=args.auto_trace, + trace_sample_mod=args.sample_mod, + max_rows=args.max_rows, + ) + await tui.start() + try: + while True: + tui.render() + await asyncio.sleep(max(0.1, args.render_interval)) + finally: + await tui.close() + + +def _build_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser(description="ezmsg profiling TUI") + parser.add_argument("--host", default=DEFAULT_HOST, help="GraphServer host") + parser.add_argument( + "--port", + type=int, + default=GRAPHSERVER_PORT_DEFAULT, + help="GraphServer port", + ) + parser.add_argument( + "--auto-start", + action="store_true", + help="Allow GraphContext to auto-start GraphServer if unavailable", + ) + parser.add_argument( + "--snapshot-interval", + type=float, + default=1.0, + help="Seconds between snapshot refreshes", + ) + parser.add_argument( + "--trace-interval", + type=float, + default=0.02, + help="Seconds between GraphServer trace stream batches", + ) + parser.add_argument( + "--max-samples", + type=int, + default=5000, + help="Max samples per process per streamed batch", + ) + parser.add_argument( + "--sample-mod", + type=int, + default=1, + help="Trace sampling divisor when auto-enabling trace", + ) + parser.add_argument( + "--render-interval", + type=float, + default=0.5, + help="Seconds between TUI redraws", + ) + parser.add_argument( + "--max-rows", + type=int, + default=30, + help="Max publisher/subscriber rows to render per table", + ) + parser.add_argument( + "--no-auto-trace", + action="store_true", + help="Do not auto-enable trace mode on discovered processes", + ) + return parser + + +def main() -> None: + parser = _build_parser() + args = parser.parse_args() + args.auto_trace = not args.no_auto_trace + asyncio.run(_run_tui(args)) + + +if __name__ == "__main__": + main() diff --git a/examples/settings_tui.py b/examples/settings_tui.py new file mode 100644 index 00000000..83acf623 --- /dev/null +++ b/examples/settings_tui.py @@ -0,0 +1,389 @@ +#!/usr/bin/env python3 +""" +Simple settings TUI for ezmsg GraphServer. + +Features: +- Live settings view (push updates via GraphContext.subscribe_settings_events) +- Inspect component metadata and current settings snapshot +- Publish patched settings to components with dynamic INPUT_SETTINGS + +Usage: + .venv/bin/python examples/settings_tui.py --host 127.0.0.1 --port 25978 + +Commands: + help + refresh + inspect + set {"field": 123, "nested": {"gain": 0.5}} + quit + +Notes: +- Updates are sent over normal pub/sub to the component's INPUT_SETTINGS topic. +- For safe updates, the script expects pickled current settings to be available + and unpickleable in this environment. +""" + +from __future__ import annotations + +import argparse +import asyncio +import contextlib +import json +import pickle +from dataclasses import dataclass, is_dataclass, replace +from typing import Any + +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import ( + ComponentMetadataType, + GraphMetadata, + SettingsChangedEvent, + SettingsSnapshotValue, +) +from ezmsg.core.netprotocol import DEFAULT_HOST, GRAPHSERVER_PORT_DEFAULT +from ezmsg.core.pubclient import Publisher + + +def _truncate(text: str, width: int) -> str: + if width <= 3: + return text[:width] + if len(text) <= width: + return text + return text[: width - 3] + "..." + + +def _format_settings(value: SettingsSnapshotValue | None, width: int = 72) -> str: + if value is None: + return "-" + return _truncate(repr(value.repr_value), width) + + +def _deep_merge_dict(base: dict[str, Any], patch: dict[str, Any]) -> dict[str, Any]: + merged = dict(base) + for key, patch_value in patch.items(): + base_value = merged.get(key) + if isinstance(base_value, dict) and isinstance(patch_value, dict): + merged[key] = _deep_merge_dict(base_value, patch_value) + else: + merged[key] = patch_value + return merged + + +def _patch_dataclass(obj: Any, patch: dict[str, Any]) -> Any: + updates: dict[str, Any] = {} + for key, patch_value in patch.items(): + if not hasattr(obj, key): + raise KeyError(f"Settings object has no field '{key}'") + current = getattr(obj, key) + if is_dataclass(current) and isinstance(patch_value, dict): + updates[key] = _patch_dataclass(current, patch_value) + elif isinstance(current, dict) and isinstance(patch_value, dict): + updates[key] = _deep_merge_dict(current, patch_value) + else: + updates[key] = patch_value + return replace(obj, **updates) + + +def _patch_value(value: Any, patch: dict[str, Any]) -> Any: + if is_dataclass(value): + return _patch_dataclass(value, patch) + if isinstance(value, dict): + return _deep_merge_dict(value, patch) + raise TypeError( + f"Cannot patch settings value of type {type(value).__name__}. " + "Only dataclass/dict settings are supported by this script." + ) + + +def _components_from_metadata( + metadata: GraphMetadata | None, +) -> dict[str, ComponentMetadataType]: + if metadata is None: + return {} + return dict(metadata.components) + + +@dataclass +class ComponentRow: + address: str + name: str + component_type: str + settings_type: str + dynamic_enabled: bool + input_topic: str | None + + +class SettingsTUI: + def __init__(self, ctx: GraphContext): + self.ctx = ctx + self.settings: dict[str, SettingsSnapshotValue] = {} + self.components: dict[str, ComponentRow] = {} + self.row_addresses: list[str] = [] + self.last_seq = 0 + self.publishers: dict[str, Publisher] = {} + self._event_queue: asyncio.Queue[SettingsChangedEvent] = asyncio.Queue() + self._watch_task: asyncio.Task[None] | None = None + + async def initialize(self) -> None: + await self.refresh() + events = await self.ctx.settings_events(after_seq=0) + for event in events: + self.settings[event.component_address] = event.value + self.last_seq = max(self.last_seq, event.seq) + self._watch_task = asyncio.create_task(self._watch_settings_events()) + + async def close(self) -> None: + if self._watch_task is not None: + self._watch_task.cancel() + with contextlib.suppress(asyncio.CancelledError): + await self._watch_task + + async def _watch_settings_events(self) -> None: + async for event in self.ctx.subscribe_settings_events(after_seq=self.last_seq): + await self._event_queue.put(event) + + async def refresh(self) -> None: + snapshot = await self.ctx.snapshot() + settings = await self.ctx.settings_snapshot() + + components: dict[str, ComponentRow] = {} + for session in snapshot.sessions.values(): + for address, comp in _components_from_metadata(session.metadata).items(): + components[address] = ComponentRow( + address=address, + name=comp.name, + component_type=comp.component_type, + settings_type=comp.settings_type, + dynamic_enabled=comp.dynamic_settings.enabled, + input_topic=comp.dynamic_settings.input_topic, + ) + + self.components = components + self.settings = settings + + async def drain_events(self) -> int: + count = 0 + while True: + try: + event = self._event_queue.get_nowait() + except asyncio.QueueEmpty: + break + self.settings[event.component_address] = event.value + self.last_seq = max(self.last_seq, event.seq) + count += 1 + return count + + def render(self, pending_updates: int = 0) -> None: + print("\x1bc", end="") + print("ezmsg settings tui") + print( + "Commands: help, refresh, inspect , " + "set , quit" + ) + if pending_updates > 0: + print(f"Applied {pending_updates} new settings event(s).") + + all_addresses = sorted(set(self.settings) | set(self.components)) + self.row_addresses = all_addresses + + header = ( + f"{'Row':<4} {'Component':<36} {'Dyn':<4} " + f"{'INPUT_SETTINGS Topic':<42} {'Current Settings':<72}" + ) + print() + print(header) + print("-" * len(header)) + + for idx, address in enumerate(all_addresses, start=1): + comp = self.components.get(address) + settings = self.settings.get(address) + + dynamic = "yes" if comp is not None and comp.dynamic_enabled else "no" + input_topic = ( + comp.input_topic if comp is not None and comp.input_topic is not None else "-" + ) + print( + f"{idx:<4} " + f"{_truncate(address, 36):<36} " + f"{dynamic:<4} " + f"{_truncate(input_topic, 42):<42} " + f"{_format_settings(settings):<72}" + ) + + def resolve_target(self, token: str) -> str: + if token.isdigit(): + idx = int(token) - 1 + if idx < 0 or idx >= len(self.row_addresses): + raise ValueError(f"Row index out of range: {token}") + return self.row_addresses[idx] + return token + + async def inspect(self, token: str) -> None: + address = self.resolve_target(token) + comp = self.components.get(address) + settings = self.settings.get(address) + print("\n--- inspect ---") + print(f"address: {address}") + if comp is None: + print("metadata: ") + else: + print(f"name: {comp.name}") + print(f"component_type: {comp.component_type}") + print(f"settings_type: {comp.settings_type}") + print(f"dynamic_settings.enabled: {comp.dynamic_enabled}") + print(f"dynamic_settings.input_topic: {comp.input_topic}") + if settings is None: + print("current_settings: ") + else: + print(f"repr: {settings.repr_value!r}") + print(f"has_pickled_payload: {settings.serialized is not None}") + if settings.serialized is not None: + try: + obj = pickle.loads(settings.serialized) + print(f"unpickled_type: {type(obj).__module__}.{type(obj).__name__}") + except Exception as exc: + print(f"unpickled_type: ") + + async def set_settings(self, token: str, patch: dict[str, Any]) -> str: + address = self.resolve_target(token) + comp = self.components.get(address) + if comp is None: + raise ValueError(f"No component metadata available for '{address}'") + if not comp.dynamic_enabled or comp.input_topic is None: + raise ValueError( + f"Component '{address}' is not dynamic-settings enabled or has no INPUT_SETTINGS topic" + ) + + current = self.settings.get(address) + if current is None: + raise ValueError(f"No current settings snapshot for '{address}'") + if current.serialized is None: + raise ValueError( + f"No serialized settings for '{address}'. Cannot safely build updated object." + ) + + try: + current_obj = pickle.loads(current.serialized) + except Exception as exc: + raise ValueError( + f"Could not unpickle current settings for '{address}': {exc}" + ) from exc + + updated_obj = _patch_value(current_obj, patch) + publisher = self.publishers.get(comp.input_topic) + if publisher is None: + publisher = await self.ctx.publisher(comp.input_topic) + self.publishers[comp.input_topic] = publisher + + await publisher.broadcast(updated_obj) + return f"Published settings update to {comp.input_topic}" + + +def _parse_patch(json_text: str) -> dict[str, Any]: + try: + patch = json.loads(json_text) + except json.JSONDecodeError as exc: + raise ValueError(f"Invalid JSON patch: {exc}") from exc + if not isinstance(patch, dict): + raise ValueError("Patch must be a JSON object") + return patch + + +def _parse_address(host: str, port: int) -> tuple[str, int]: + return (host, port) + + +async def _run_tui(host: str, port: int, auto_start: bool) -> None: + address = _parse_address(host, port) + + async with GraphContext(address, auto_start=auto_start) as ctx: + tui = SettingsTUI(ctx) + await tui.initialize() + try: + while True: + pending = await tui.drain_events() + tui.render(pending_updates=pending) + cmdline = (await asyncio.to_thread(input, "\nsettings-tui> ")).strip() + if not cmdline: + continue + + cmd, *rest = cmdline.split(" ", 1) + if cmd in {"q", "quit", "exit"}: + break + + if cmd in {"h", "help"}: + print( + "\nhelp:\n" + " refresh\n" + " inspect \n" + " set \n" + " quit\n" + ) + await asyncio.to_thread(input, "Press Enter to continue...") + continue + + if cmd == "refresh": + await tui.refresh() + continue + + if cmd == "inspect": + if not rest: + print("Usage: inspect ") + else: + await tui.inspect(rest[0].strip()) + await asyncio.to_thread(input, "Press Enter to continue...") + continue + + if cmd == "set": + if not rest: + print("Usage: set ") + await asyncio.to_thread(input, "Press Enter to continue...") + continue + + target_and_patch = rest[0].strip() + if " " not in target_and_patch: + print("Usage: set ") + await asyncio.to_thread(input, "Press Enter to continue...") + continue + + target, patch_text = target_and_patch.split(" ", 1) + try: + patch = _parse_patch(patch_text.strip()) + result = await tui.set_settings(target.strip(), patch) + print(result) + except Exception as exc: + print(f"set failed: {exc}") + await asyncio.to_thread(input, "Press Enter to continue...") + continue + + print(f"Unknown command: {cmd}") + await asyncio.to_thread(input, "Press Enter to continue...") + finally: + await tui.close() + + +def _build_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser(description="ezmsg settings TUI") + parser.add_argument("--host", default=DEFAULT_HOST, help="GraphServer host") + parser.add_argument( + "--port", + type=int, + default=GRAPHSERVER_PORT_DEFAULT, + help="GraphServer port", + ) + parser.add_argument( + "--auto-start", + action="store_true", + help="Allow GraphContext to auto-start GraphServer if unavailable", + ) + return parser + + +def main() -> None: + parser = _build_parser() + args = parser.parse_args() + asyncio.run(_run_tui(args.host, args.port, args.auto_start)) + + +if __name__ == "__main__": + main() diff --git a/examples/topology_tui.py b/examples/topology_tui.py new file mode 100644 index 00000000..56876a34 --- /dev/null +++ b/examples/topology_tui.py @@ -0,0 +1,285 @@ +#!/usr/bin/env python3 +""" +Simple live topology TUI for ezmsg GraphServer. + +Features: +- Push-based topology event subscription +- Live graph summary (nodes/edges/sessions/processes) +- Process ownership view +- Current edge list +- Recent topology event log + +Usage: + PYTHONPATH=src .venv/bin/python examples/topology_tui.py --host 127.0.0.1 --port 25978 +""" + +from __future__ import annotations + +import argparse +import asyncio +import contextlib +import time +from collections import deque + +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import GraphSnapshot, TopologyChangedEvent +from ezmsg.core.netprotocol import DEFAULT_HOST, GRAPHSERVER_PORT_DEFAULT + + +def _truncate(text: object, width: int) -> str: + text = str(text) + if width <= 3: + return text[:width] + if len(text) <= width: + return text + return text[: width - 3] + "..." + + +def _fmt_age(age_s: float) -> str: + return f"{age_s:0.2f}s" + + +def _flatten_edges(snapshot: GraphSnapshot) -> list[tuple[str, str]]: + edges: list[tuple[str, str]] = [] + for src, destinations in snapshot.graph.items(): + for dst in destinations: + edges.append((src, dst)) + edges.sort(key=lambda edge: (edge[0], edge[1])) + return edges + + +class TopologyTUI: + def __init__( + self, + ctx: GraphContext, + *, + snapshot_interval: float, + render_interval: float, + max_edges: int, + max_events: int, + max_processes: int, + ) -> None: + self.ctx = ctx + self.snapshot_interval = max(0.2, snapshot_interval) + self.render_interval = max(0.1, render_interval) + self.max_edges = max(10, max_edges) + self.max_events = max(10, max_events) + self.max_processes = max(5, max_processes) + + self.snapshot: GraphSnapshot | None = None + self.last_snapshot_time: float | None = None + self._events: deque[TopologyChangedEvent] = deque(maxlen=self.max_events) + self._event_queue: asyncio.Queue[TopologyChangedEvent] = asyncio.Queue() + + self._watch_task: asyncio.Task[None] | None = None + + async def start(self) -> None: + await self._refresh_snapshot() + self._watch_task = asyncio.create_task(self._watch_topology_events()) + + async def close(self) -> None: + if self._watch_task is not None: + self._watch_task.cancel() + with contextlib.suppress(asyncio.CancelledError): + await self._watch_task + + async def _watch_topology_events(self) -> None: + after_seq = 0 + async for event in self.ctx.subscribe_topology_events(after_seq=after_seq): + after_seq = event.seq + await self._event_queue.put(event) + + async def _refresh_snapshot(self) -> None: + self.snapshot = await self.ctx.snapshot() + self.last_snapshot_time = time.time() + + async def update(self) -> int: + """ + Drain queued topology events and refresh snapshot if needed. + + Returns: + Number of drained events. + """ + drained = 0 + refresh_requested = False + + while True: + try: + event = self._event_queue.get_nowait() + except asyncio.QueueEmpty: + break + self._events.append(event) + refresh_requested = True + drained += 1 + + if self.last_snapshot_time is None: + await self._refresh_snapshot() + elif refresh_requested or (time.time() - self.last_snapshot_time) >= self.snapshot_interval: + await self._refresh_snapshot() + + return drained + + def render(self, drained_events: int) -> None: + print("\x1bc", end="") + print("ezmsg topology tui") + print("Ctrl-C to quit") + print( + f"snapshot_interval={self.snapshot_interval:.2f}s " + f"render_interval={self.render_interval:.2f}s" + ) + if self.last_snapshot_time is not None: + print(f"snapshot_age={_fmt_age(max(0.0, time.time() - self.last_snapshot_time))}") + if drained_events > 0: + print(f"applied_events={drained_events}") + + snapshot = self.snapshot + if snapshot is None: + print("\n") + return + + edges = _flatten_edges(snapshot) + node_names = set(snapshot.graph.keys()) + for _, dst in edges: + node_names.add(dst) + + print( + "\nsummary: " + f"nodes={len(node_names)} edges={len(edges)} " + f"sessions={len(snapshot.sessions)} processes={len(snapshot.processes)}" + ) + + print("\nprocesses") + proc_header = f"{'Process':<30} {'PID':>8} {'Host':<24} {'Units':<80}" + print(proc_header) + print("-" * len(proc_header)) + if not snapshot.processes: + print("") + else: + process_items = sorted(snapshot.processes.values(), key=lambda p: p.process_id) + for proc in process_items[: self.max_processes]: + units = ", ".join(proc.units) if proc.units else "-" + print( + f"{_truncate(proc.process_id, 30):<30} " + f"{str(proc.pid) if proc.pid is not None else '-':>8} " + f"{_truncate(proc.host if proc.host is not None else '-', 24):<24} " + f"{_truncate(units, 80):<80}" + ) + if len(process_items) > self.max_processes: + print(f"... {len(process_items) - self.max_processes} more process rows") + + print("\nedges") + edge_header = f"{'From':<48} {'To':<48}" + print(edge_header) + print("-" * len(edge_header)) + if not edges: + print("") + else: + for src, dst in edges[: self.max_edges]: + print(f"{_truncate(src, 48):<48} {_truncate(dst, 48):<48}") + if len(edges) > self.max_edges: + print(f"... {len(edges) - self.max_edges} more edges") + + print("\nrecent topology events") + event_header = ( + f"{'Seq':>6} {'Type':<15} {'Age':>8} {'Topics':<44} " + f"{'Source Session':<38} {'Source Process':<30}" + ) + print(event_header) + print("-" * len(event_header)) + if not self._events: + print("") + else: + now = time.time() + for event in reversed(self._events): + topics = ", ".join(event.changed_topics) if event.changed_topics else "-" + print( + f"{event.seq:>6} " + f"{_truncate(event.event_type.value, 15):<15} " + f"{_fmt_age(max(0.0, now - event.timestamp)):>8} " + f"{_truncate(topics, 44):<44} " + f"{_truncate(event.source_session_id or '-', 38):<38} " + f"{_truncate(event.source_process_id or '-', 30):<30}" + ) + + +def _parse_address(host: str, port: int) -> tuple[str, int]: + return (host, port) + + +async def _run_tui(args: argparse.Namespace) -> None: + async with GraphContext( + _parse_address(args.host, args.port), auto_start=args.auto_start + ) as ctx: + tui = TopologyTUI( + ctx, + snapshot_interval=args.snapshot_interval, + render_interval=args.render_interval, + max_edges=args.max_edges, + max_events=args.max_events, + max_processes=args.max_processes, + ) + await tui.start() + try: + while True: + drained = await tui.update() + tui.render(drained_events=drained) + await asyncio.sleep(tui.render_interval) + finally: + await tui.close() + + +def _build_parser() -> argparse.ArgumentParser: + parser = argparse.ArgumentParser(description="ezmsg topology TUI") + parser.add_argument("--host", default=DEFAULT_HOST, help="GraphServer host") + parser.add_argument( + "--port", + type=int, + default=GRAPHSERVER_PORT_DEFAULT, + help="GraphServer port", + ) + parser.add_argument( + "--auto-start", + action="store_true", + help="Allow GraphContext to auto-start GraphServer if unavailable", + ) + parser.add_argument( + "--snapshot-interval", + type=float, + default=1.0, + help="Seconds between forced snapshot refreshes", + ) + parser.add_argument( + "--render-interval", + type=float, + default=0.5, + help="Seconds between screen redraws", + ) + parser.add_argument( + "--max-edges", + type=int, + default=50, + help="Max edge rows to render", + ) + parser.add_argument( + "--max-events", + type=int, + default=25, + help="Max recent topology events to retain/render", + ) + parser.add_argument( + "--max-processes", + type=int, + default=20, + help="Max process rows to render", + ) + return parser + + +def main() -> None: + parser = _build_parser() + args = parser.parse_args() + asyncio.run(_run_tui(args)) + + +if __name__ == "__main__": + main() diff --git a/scripts/perf_ab.py b/scripts/perf_ab.py new file mode 100644 index 00000000..5e7a8b8d --- /dev/null +++ b/scripts/perf_ab.py @@ -0,0 +1,5 @@ +from ezmsg.util.perf.ab import main + + +if __name__ == "__main__": + main() diff --git a/src/ezmsg/core/__init__.py b/src/ezmsg/core/__init__.py index bc56a3b9..aaf36899 100644 --- a/src/ezmsg/core/__init__.py +++ b/src/ezmsg/core/__init__.py @@ -14,6 +14,11 @@ "Settings", "Collection", "NetworkDefinition", + "Topic", + "InputTopic", + "OutputTopic", + "InputRelay", + "OutputRelay", "InputStream", "OutputStream", "Unit", @@ -44,7 +49,15 @@ from .settings import Settings from .collection import Collection, NetworkDefinition from .unit import Unit, task, publisher, subscriber, main, timeit, process, thread -from .stream import InputStream, OutputStream +from .stream import ( + Topic, + InputTopic, + OutputTopic, + InputRelay, + OutputRelay, + InputStream, + OutputStream, +) from .backend import run, GraphRunner, GraphRunnerStartError from .backendprocess import Complete, NormalTermination from .graphserver import GraphServer diff --git a/src/ezmsg/core/backend.py b/src/ezmsg/core/backend.py index 7a0aba1d..7a06fbd0 100644 --- a/src/ezmsg/core/backend.py +++ b/src/ezmsg/core/backend.py @@ -2,9 +2,12 @@ from collections.abc import Callable, Mapping, Iterable from collections.abc import Collection as AbstractCollection import enum +import inspect import logging import os +import pickle import signal +from dataclasses import dataclass from threading import BrokenBarrierError from multiprocessing import Event, Barrier from multiprocessing.synchronize import Event as EventType @@ -16,8 +19,44 @@ from .collection import Collection, NetworkDefinition from .component import Component -from .stream import Stream -from .unit import Unit, PROCESS_ATTR +from .stream import ( + Stream, + InputStream, + OutputStream, + Topic, + InputTopic, + OutputTopic, + InputRelay, + OutputRelay, +) +from .unit import Unit, PROCESS_ATTR, SUBSCRIBES_ATTR, PUBLISHES_ATTR +from .settings import Settings +from .graphmeta import ( + CollectionMetadata, + ComponentMetadata, + ComponentMetadataType, + DynamicSettingsMetadata, + InputRelayMetadata, + InputStreamMetadata, + InputTopicMetadata, + OutputRelayMetadata, + OutputStreamMetadata, + OutputTopicMetadata, + RelayMetadataType, + StreamMetadataType, + StreamMetadata, + TopicMetadata, + TopicMetadataType, + TaskMetadata, + GraphMetadata, + UnitMetadata, +) +from .relay import _CollectionRelayUnit, _RelaySettings +from .settingsmeta import ( + settings_repr_value, + settings_schema_from_type, + settings_schema_from_value, +) from .graphserver import GraphService from .graphcontext import GraphContext @@ -33,6 +72,31 @@ logger = logging.getLogger("ezmsg") +def crawl_components( + component: Component, + callback: Callable[[Component], None] | None = None, +) -> list[Component]: + search: list[Component] = [component] + out: list[Component] = [] + while len(search): + comp = search.pop() + out.append(comp) + search += list(comp.components.values()) + if callback is not None: + callback(comp) + return out + + +@dataclass +class _RelayBinding: + kind: str # "input" or "output" + endpoint_topic: str + relay_in_topic: str + relay_out_topic: str + endpoint: InputRelay | OutputRelay + relay_unit: _CollectionRelayUnit + + class ExecutionContext: _process_units: list[list[Unit]] _processes: list[BackendProcess] | None @@ -95,49 +159,141 @@ def setup( start_participant: bool = False, ) -> "ExecutionContext | None": graph_connections: list[tuple[str, str]] = [] + relay_bindings: dict[str, _RelayBinding] = {} for name, component in components.items(): component._set_name(name) component._set_location([root_name] if root_name is not None else []) + def normalize_topic(endpoint: Stream | str | enum.Enum, where: str) -> str: + if isinstance(endpoint, Stream): + return endpoint.address + if isinstance(endpoint, enum.Enum): + return endpoint.name + if isinstance(endpoint, str): + return endpoint + raise TypeError( + f"Invalid endpoint type in {where}: {type(endpoint)}. " + "Expected Stream, str, or Enum." + ) + if connections is not None: for from_topic, to_topic in connections: - if isinstance(from_topic, Stream): - from_topic = from_topic.address - if isinstance(to_topic, Stream): - to_topic = to_topic.address - if isinstance(to_topic, enum.Enum): - to_topic = to_topic.name - if isinstance(from_topic, enum.Enum): - from_topic = from_topic.name - graph_connections.append((from_topic, to_topic)) - - def crawl_components( - component: Component, callback: Callable[[Component], None] - ) -> None: - search: list[Component] = [component] - while len(search): - comp = search.pop() - search += list(comp.components.values()) - callback(comp) + graph_connections.append( + ( + normalize_topic(from_topic, "connections"), + normalize_topic(to_topic, "connections"), + ) + ) + def input_relay_settings(relay: InputRelay) -> _RelaySettings: + return _RelaySettings( + leaky=relay.leaky, + max_queue=relay.max_queue, + copy_on_forward=relay.copy_on_forward, + ) + + def output_relay_settings(relay: OutputRelay) -> _RelaySettings: + return _RelaySettings( + host=relay.host, + port=relay.port, + num_buffers=relay.num_buffers, + buf_size=relay.buf_size, + force_tcp=relay.force_tcp, + copy_on_forward=relay.copy_on_forward, + ) + + def add_collection_relay_units(comp: Component) -> None: + if not isinstance(comp, Collection): + return + + for endpoint_name, endpoint in comp.streams.items(): + if isinstance(endpoint, InputRelay): + relay_name = f"__relay_in_{endpoint_name}" + if relay_name in comp.components: + raise ValueError( + f"{comp.address} already defines component '{relay_name}'." + ) + + relay_unit = _CollectionRelayUnit(input_relay_settings(endpoint)) + relay_unit._set_name(relay_name) + relay_unit._set_location(comp.location + [comp.name]) + comp.components[relay_name] = relay_unit + setattr(comp, relay_name, relay_unit) + + relay_bindings[endpoint.address] = _RelayBinding( + kind="input", + endpoint_topic=endpoint.address, + relay_in_topic=relay_unit.INPUT.address, + relay_out_topic=relay_unit.OUTPUT.address, + endpoint=endpoint, + relay_unit=relay_unit, + ) + + elif isinstance(endpoint, OutputRelay): + relay_name = f"__relay_out_{endpoint_name}" + if relay_name in comp.components: + raise ValueError( + f"{comp.address} already defines component '{relay_name}'." + ) + + relay_unit = _CollectionRelayUnit(output_relay_settings(endpoint)) + relay_unit._set_name(relay_name) + relay_unit._set_location(comp.location + [comp.name]) + comp.components[relay_name] = relay_unit + setattr(comp, relay_name, relay_unit) + + relay_bindings[endpoint.address] = _RelayBinding( + kind="output", + endpoint_topic=endpoint.address, + relay_in_topic=relay_unit.INPUT.address, + relay_out_topic=relay_unit.OUTPUT.address, + endpoint=endpoint, + relay_unit=relay_unit, + ) + + for component in components.values(): + if isinstance(component, Collection): + crawl_components(component, add_collection_relay_units) def gather_edges(comp: Component): if isinstance(comp, Collection): for from_stream, to_stream in comp.network(): - if isinstance(from_stream, Stream): - from_stream = from_stream.address - if isinstance(to_stream, Stream): - to_stream = to_stream.address - if isinstance(to_stream, enum.Enum): - to_stream = to_stream.name - if isinstance(from_stream, enum.Enum): - from_stream = from_stream.name - graph_connections.append((from_stream, to_stream)) + graph_connections.append( + ( + normalize_topic(from_stream, f"{comp.address}.network"), + normalize_topic(to_stream, f"{comp.address}.network"), + ) + ) for component in components.values(): if isinstance(component, Collection): crawl_components(component, gather_edges) + if relay_bindings: + rewritten_connections: list[tuple[str, str]] = [] + for from_topic, to_topic in graph_connections: + to_binding = relay_bindings.get(to_topic, None) + if to_binding is not None and to_binding.kind == "output": + to_topic = to_binding.relay_in_topic + + from_binding = relay_bindings.get(from_topic, None) + if from_binding is not None and from_binding.kind == "input": + from_topic = from_binding.relay_out_topic + + rewritten_connections.append((from_topic, to_topic)) + + for binding in relay_bindings.values(): + if binding.kind == "input": + rewritten_connections.append( + (binding.endpoint_topic, binding.relay_in_topic) + ) + else: + rewritten_connections.append( + (binding.relay_out_topic, binding.endpoint_topic) + ) + + graph_connections = rewritten_connections + processes = collect_processes(components.values(), process_components) for component in components.values(): @@ -149,6 +305,14 @@ def configure_collections(comp: Component): crawl_components(component, configure_collections) + for binding in relay_bindings.values(): + if isinstance(binding.endpoint, InputRelay): + binding.relay_unit.apply_settings(input_relay_settings(binding.endpoint)) + elif isinstance(binding.endpoint, OutputRelay): + binding.relay_unit.apply_settings( + output_relay_settings(binding.endpoint) + ) + if force_single_process: processes = [[u for pu in processes for u in pu]] @@ -255,6 +419,206 @@ def processes(self) -> list[BackendProcess]: def running(self) -> bool: return self._started + def _type_name(self, tp: type) -> str: + return f"{tp.__module__}.{tp.__qualname__}" + + def _stream_type_name(self, stream_type: object) -> str: + if inspect.isclass(stream_type): + return self._type_name(stream_type) + return repr(stream_type) + + def _settings_repr(self, value: object) -> dict[str, object] | str: + return settings_repr_value(value) + + def _settings_snapshot(self, value: object) -> tuple[bytes | None, dict[str, object] | str]: + try: + pickled = pickle.dumps(value) + except Exception as exc: + logger.warning(f"Could not pickle settings for metadata: {exc}") + pickled = None + return pickled, self._settings_repr(value) + + def _component_metadata(self) -> GraphMetadata: + components: dict[str, ComponentMetadataType] = {} + + for root in self._components.values(): + for comp in crawl_components(root): + is_collection = isinstance(comp, Collection) + input_settings = comp.streams.get("INPUT_SETTINGS") + dynamic_settings = DynamicSettingsMetadata( + enabled=isinstance(input_settings, InputStream), + input_topic=( + input_settings.address + if isinstance(input_settings, InputStream) + else None + ), + settings_type=( + self._stream_type_name(input_settings.msg_type) + if isinstance(input_settings, InputStream) + else None + ), + ) + + stream_entries: dict[str, StreamMetadataType] = {} + topic_entries: dict[str, TopicMetadataType] = {} + relay_entries: dict[str, RelayMetadataType] = {} + for stream_name, stream in comp.streams.items(): + msg_type = self._stream_type_name(stream.msg_type) + if isinstance(stream, InputRelay): + relay_entries[stream_name] = InputRelayMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + leaky=stream.leaky, + max_queue=stream.max_queue, + copy_on_forward=stream.copy_on_forward, + ) + elif isinstance(stream, OutputRelay): + relay_entries[stream_name] = OutputRelayMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + host=stream.host, + port=stream.port, + num_buffers=stream.num_buffers, + buf_size=stream.buf_size, + force_tcp=stream.force_tcp, + copy_on_forward=stream.copy_on_forward, + ) + elif isinstance(stream, InputTopic): + topic_entries[stream_name] = InputTopicMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + ) + elif isinstance(stream, OutputTopic): + topic_entries[stream_name] = OutputTopicMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + ) + elif isinstance(stream, Topic): + topic_entries[stream_name] = TopicMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + ) + elif isinstance(stream, InputStream): + if is_collection: + topic_entries[stream_name] = InputTopicMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + ) + else: + stream_entries[stream_name] = InputStreamMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + leaky=stream.leaky, + max_queue=stream.max_queue, + ) + elif isinstance(stream, OutputStream): + if is_collection: + topic_entries[stream_name] = OutputTopicMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + ) + else: + stream_entries[stream_name] = OutputStreamMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + host=stream.host, + port=stream.port, + num_buffers=stream.num_buffers, + buf_size=stream.buf_size, + force_tcp=stream.force_tcp, + ) + else: + if is_collection: + topic_entries[stream_name] = TopicMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + ) + else: + stream_entries[stream_name] = StreamMetadata( + name=stream_name, + address=stream.address, + msg_type=msg_type, + ) + + task_entries: list[TaskMetadata] = [] + for task_name, task in comp.tasks.items(): + task_entry = TaskMetadata(name=task_name) + + if hasattr(task, SUBSCRIBES_ATTR): + sub_stream = getattr(task, SUBSCRIBES_ATTR) + if hasattr(sub_stream, "name") and sub_stream.name in comp.streams: + task_entry.subscribes = comp.streams[sub_stream.name].address + + if hasattr(task, PUBLISHES_ATTR): + pub_streams = getattr(task, PUBLISHES_ATTR) + task_entry.publishes = [ + comp.streams[stream.name].address + for stream in pub_streams + if hasattr(stream, "name") and stream.name in comp.streams + ] + + task_entries.append(task_entry) + + settings_type = getattr(comp.__class__, "__settings_type__", Settings) + settings_type_name = ( + self._type_name(settings_type) + if inspect.isclass(settings_type) + else repr(settings_type) + ) + settings_schema = ( + settings_schema_from_value(comp.SETTINGS) + if comp.SETTINGS is not None + else settings_schema_from_type(settings_type) + ) + + component_common = dict( + address=comp.address, + name=comp.name, + component_type=self._type_name(comp.__class__), + settings_type=settings_type_name, + initial_settings=self._settings_snapshot(comp.SETTINGS), + dynamic_settings=dynamic_settings, + settings_schema=settings_schema, + ) + + metadata_entry: ComponentMetadataType + if isinstance(comp, Collection): + metadata_entry = CollectionMetadata( + **component_common, + topics=topic_entries, + relays=relay_entries, + children=sorted( + child.address for child in comp.components.values() + ), + ) + elif isinstance(comp, Unit): + metadata_entry = UnitMetadata( + **component_common, + streams=stream_entries, + tasks=sorted(task_entries, key=lambda task: task.name), + main=comp.main.__name__ if comp.main is not None else None, + threads=sorted(comp.threads.keys()), + ) + else: + metadata_entry = ComponentMetadata(**component_common) + components[comp.address] = metadata_entry + + return GraphMetadata( + schema_version=1, + root_name=self._root_name, + components={address: components[address] for address in sorted(components)}, + ) + def start(self) -> None: if self._started: raise RuntimeError("GraphRunner is already running") @@ -360,6 +724,15 @@ async def setup_graph() -> None: asyncio.run_coroutine_threadsafe(setup_graph(), self._loop).result() + metadata = self._component_metadata() + + async def register_graph_metadata() -> None: + await graph_context.register_metadata(metadata) + + asyncio.run_coroutine_threadsafe( + register_graph_metadata(), self._loop + ).result() + if len(self._execution_context.processes) > 1: logger.info( f"Running in {len(self._execution_context.processes)} processes." diff --git a/src/ezmsg/core/backendprocess.py b/src/ezmsg/core/backendprocess.py index 8ef22a72..0216cab9 100644 --- a/src/ezmsg/core/backendprocess.py +++ b/src/ezmsg/core/backendprocess.py @@ -3,22 +3,31 @@ import logging import inspect import os -import time +import pickle import traceback import threading import weakref +from copy import deepcopy from abc import abstractmethod -from dataclasses import dataclass +from dataclasses import dataclass, fields as dataclass_fields, is_dataclass, replace from collections import defaultdict -from collections.abc import Callable, Coroutine, Generator, Sequence +from collections.abc import ( + AsyncGenerator, + Awaitable, + Callable, + Coroutine, + Generator, + Mapping, + Sequence, +) from functools import wraps, partial from concurrent.futures import ThreadPoolExecutor from concurrent.futures.thread import _worker from multiprocessing import Process from multiprocessing.synchronize import Event as EventType from multiprocessing.synchronize import Barrier as BarrierType -from contextlib import suppress, contextmanager +from contextlib import suppress, contextmanager, asynccontextmanager from concurrent.futures import TimeoutError from typing import Any @@ -26,9 +35,24 @@ from .unit import Unit, TIMEIT_ATTR, SUBSCRIBES_ATTR from .graphcontext import GraphContext +from .graphmeta import ( + ProcessControlErrorCode, + ProcessControlOperation, + ProcessControlRequest, + ProcessControlResponse, + SettingsFieldUpdateRequest, + SettingsSnapshotValue, +) +from .profiling import PROFILES, PROFILE_TIME +from .processclient import ProcessControlClient from .pubclient import Publisher from .subclient import Subscriber from .netprotocol import AddressType +from .settingsmeta import ( + settings_repr_value, + settings_schema_from_value, + settings_structured_value, +) logger = logging.getLogger("ezmsg") @@ -220,12 +244,188 @@ class DefaultBackendProcess(BackendProcess): pubs: dict[str, Publisher] _shutdown_errors: bool + def _settings_snapshot_value(self, value: object) -> SettingsSnapshotValue: + try: + serialized = pickle.dumps(value) + except Exception: + serialized = None + + return SettingsSnapshotValue( + serialized=serialized, + repr_value=settings_repr_value(value), + structured_value=settings_structured_value(value), + settings_schema=settings_schema_from_value(value), + ) + + def _replace_settings_field( + self, settings_value: object, field_path: str, value: object + ) -> object: + if field_path == "": + raise ValueError("field_path must not be empty") + path = field_path.split(".") + + def apply(current: object, idx: int) -> object: + field_name = path[idx] + if isinstance(current, Mapping): + if field_name not in current: + raise AttributeError( + f"Settings field '{field_name}' does not exist in mapping" + ) + if idx == len(path) - 1: + updated = dict(current) + updated[field_name] = value + return updated + patched_child = apply(current[field_name], idx + 1) + updated = dict(current) + updated[field_name] = patched_child + return updated + + if not hasattr(current, field_name): + raise AttributeError( + f"Settings field '{field_name}' does not exist on " + f"{type(current).__name__}" + ) + + if idx == len(path) - 1: + return self._patch_object_field(current, field_name, value) + + child_value = getattr(current, field_name) + patched_child = apply(child_value, idx + 1) + return self._patch_object_field(current, field_name, patched_child) + + return apply(settings_value, 0) + + def _patch_object_field( + self, obj: object, field_name: str, value: object + ) -> object: + if is_dataclass(obj): + valid_fields = {f.name for f in dataclass_fields(obj)} + if field_name not in valid_fields: + raise AttributeError( + f"Settings field '{field_name}' does not exist on " + f"{type(obj).__name__}" + ) + return replace(obj, **{field_name: value}) + + if hasattr(obj, "model_copy") and callable(getattr(obj, "model_copy")): + return obj.model_copy(update={field_name: value}) # type: ignore[attr-defined] + + if hasattr(obj, "copy") and callable(getattr(obj, "copy")): + try: + return obj.copy(update={field_name: value}) # type: ignore[attr-defined] + except Exception: + pass + + if hasattr(obj, field_name): + patched = deepcopy(obj) + setattr(patched, field_name, value) + return patched + + raise TypeError(f"Cannot patch settings object of type {type(obj).__name__}") + def process(self, loop: asyncio.AbstractEventLoop) -> None: main_func = None context = GraphContext(self.graph_address) + process_client = ProcessControlClient(self.graph_address) coro_callables: dict[str, Callable[[], Coroutine[Any, Any, None]]] = dict() + settings_input_topics: dict[str, str] = {} + current_settings: dict[str, object] = {} + control_publishers: dict[str, Publisher] = {} self._shutdown_errors = False + async def process_request_handler( + request: ProcessControlRequest, + ) -> ProcessControlResponse: + if request.operation != ProcessControlOperation.UPDATE_SETTING_FIELD.value: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=f"Unsupported process control operation: {request.operation}", + error_code=ProcessControlErrorCode.UNSUPPORTED_OPERATION, + process_id=process_client.process_id, + ) + + if request.payload is None: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error="Missing settings field update payload", + error_code=ProcessControlErrorCode.INVALID_RESPONSE, + process_id=process_client.process_id, + ) + + try: + update_obj = pickle.loads(request.payload) + if not isinstance(update_obj, SettingsFieldUpdateRequest): + raise RuntimeError( + "settings field update payload was not SettingsFieldUpdateRequest" + ) + except Exception as exc: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=f"Invalid settings field update payload: {exc}", + error_code=ProcessControlErrorCode.INVALID_RESPONSE, + process_id=process_client.process_id, + ) + + unit_address = request.unit_address + input_topic = settings_input_topics.get(unit_address) + if input_topic is None: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=( + f"Unit '{unit_address}' does not expose INPUT_SETTINGS; " + "settings field update unsupported" + ), + error_code=ProcessControlErrorCode.UNSUPPORTED_OPERATION, + process_id=process_client.process_id, + ) + + if unit_address not in current_settings: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=( + f"No current settings value tracked for unit '{unit_address}'. " + "Send a full settings object first via update_settings()." + ), + error_code=ProcessControlErrorCode.HANDLER_ERROR, + process_id=process_client.process_id, + ) + + try: + patched = self._replace_settings_field( + current_settings[unit_address], + update_obj.field_path, + update_obj.value, + ) + control_pub = control_publishers.get(input_topic) + if control_pub is None: + control_pub = await context.publisher(input_topic) + control_publishers[input_topic] = control_pub + await control_pub.broadcast(patched) + current_settings[unit_address] = patched + except Exception as exc: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=f"Failed to patch settings field: {exc}", + error_code=ProcessControlErrorCode.HANDLER_ERROR, + process_id=process_client.process_id, + ) + + result_value = self._settings_snapshot_value(patched) + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + payload=pickle.dumps(result_value), + process_id=process_client.process_id, + ) + + process_client.set_request_handler(process_request_handler) + try: self.pubs = dict() @@ -259,6 +459,8 @@ async def setup_state(): main_func = None for unit in self.units: + if unit.SETTINGS is not None: + current_settings[unit.address] = unit.SETTINGS sub_callables: defaultdict[ str, set[Callable[..., Coroutine[Any, Any, None]]] ] = defaultdict(set) @@ -284,8 +486,32 @@ async def setup_state(): loop, ).result() task_name = f"SUBSCRIBER|{stream.address}" + report_settings_update: ( + Callable[[object], Awaitable[None]] | None + ) = None + if stream.name == "INPUT_SETTINGS": + component_address = unit.address + settings_input_topics[component_address] = stream.address + + async def report_settings_update_cb( + msg: object, + *, + _component_address: str = component_address, + ) -> None: + current_settings[_component_address] = msg + value = self._settings_snapshot_value(msg) + await process_client.report_settings_update( + component_address=_component_address, + value=value, + ) + + report_settings_update = report_settings_update_cb + coro_callables[task_name] = partial( - handle_subscriber, sub, sub_callables[stream.address] + handle_subscriber, + sub, + sub_callables[stream.address], + on_message=report_settings_update, ) elif isinstance(stream, OutputStream): @@ -299,10 +525,16 @@ async def setup_state(): buf_size=stream.buf_size, start_paused=True, force_tcp=stream.force_tcp, + allow_local=stream.allow_local, ), loop=loop, ).result() + asyncio.run_coroutine_threadsafe( + process_client.register([unit.address for unit in self.units]), + loop, + ).result() + except asyncio.CancelledError: pass @@ -407,6 +639,13 @@ async def shutdown_units() -> None: except TimeoutError: logger.warning("Timed out waiting for retry on context revert") + process_close_future = asyncio.run_coroutine_threadsafe( + process_client.close(), + loop=loop, + ) + with suppress(Exception): + process_close_future.result() + logger.debug(f"Remaining tasks in event loop = {asyncio.all_tasks(loop)}") if self.task_finished_ev is not None: @@ -439,11 +678,12 @@ async def publish(stream: Stream, obj: Any) -> None: await asyncio.sleep(0) async def perf_publish(stream: Stream, obj: Any) -> None: - start = time.perf_counter() + start = PROFILE_TIME() await publish(stream, obj) - stop = time.perf_counter() + stop = PROFILE_TIME() logger.info( - f"{task_address} send duration = " + f"{(stop - start) * 1e3:0.4f}ms" + f"{task_address} send duration = " + f"{((stop - start) / 1_000_000.0):0.4f}ms" ) pub_fn = perf_publish if hasattr(task, TIMEIT_ATTR) else publish @@ -487,8 +727,10 @@ async def wrapped_task(msg: Any = None) -> None: except Exception: logger.error(f"Exception in Task: {task_address}") logger.error(traceback.format_exc()) - if self.term_ev.is_set(): - self._shutdown_errors = True + # Any task exception should mark shutdown as unclean so + # interrupt-driven teardown can return a non-zero exit code. + # Gating this on term_ev introduces timing-dependent behavior. + self._shutdown_errors = True if strict_shutdown: raise @@ -496,7 +738,9 @@ async def wrapped_task(msg: Any = None) -> None: async def handle_subscriber( - sub: Subscriber, callables: set[Callable[..., Coroutine[Any, Any, None]]] + sub: Subscriber, + callables: set[Callable[..., Coroutine[Any, Any, None]]], + on_message: Callable[[Any], Awaitable[None]] | None = None, ): """ Handle incoming messages from a subscriber and distribute to callables. @@ -515,32 +759,47 @@ async def handle_subscriber( # Non-leaky subscribers use recv_zero_copy() to hold backpressure during # processing, which provides zero-copy performance but applies backpressure. + @asynccontextmanager + async def next_message() -> AsyncGenerator[Any, None]: + if sub.leaky: + msg = await sub.recv() + try: + yield msg + finally: + del msg + return + + async with sub.recv_zero_copy() as msg: + yield msg + while True: if not callables: sub.close() await sub.wait_closed() break - if sub.leaky: - msg = await sub.recv() + async with next_message() as msg: try: + if on_message is not None: + try: + await on_message(msg) + except Exception as exc: + logger.warning( + f"Failed to report subscriber message metadata: {exc}" + ) for callable in list(callables): try: - await callable(msg) + span_start_ns = sub.begin_profile() + try: + await callable(msg) + finally: + sub.end_profile( + span_start_ns, getattr(callable, "__name__", None) + ) except (Complete, NormalTermination): callables.remove(callable) finally: del msg - else: - async with sub.recv_zero_copy() as msg: - try: - for callable in list(callables): - try: - await callable(msg) - except (Complete, NormalTermination): - callables.remove(callable) - finally: - del msg if len(callables) > 1: await asyncio.sleep(0) diff --git a/src/ezmsg/core/collection.py b/src/ezmsg/core/collection.py index dd8f665d..1f1f51e6 100644 --- a/src/ezmsg/core/collection.py +++ b/src/ezmsg/core/collection.py @@ -2,8 +2,9 @@ from collections.abc import Collection as AbstractCollection import typing from copy import deepcopy +import warnings -from .stream import Stream +from .stream import Stream, InputStream, OutputStream from .component import ComponentMeta, Component from .settings import Settings @@ -34,6 +35,16 @@ def __init__( if isinstance(field_value, Component): field_value._set_name(field_name) cls.__components__[field_name] = field_value + elif isinstance(field_value, (InputStream, OutputStream)): + warnings.warn( + f"{name}.{field_name} uses {type(field_value).__name__} as a " + "Collection boundary endpoint. This behavior is deprecated and " + "will change in a future release. Use InputTopic / OutputTopic " + "for zero-overhead topic shortcuts, or InputRelay / OutputRelay " + "for explicit boundary republishers.", + FutureWarning, + stacklevel=2, + ) class Collection(Component, metaclass=CollectionMeta): diff --git a/src/ezmsg/core/dag.py b/src/ezmsg/core/dag.py index d0c9a72a..1ae5290e 100644 --- a/src/ezmsg/core/dag.py +++ b/src/ezmsg/core/dag.py @@ -1,6 +1,7 @@ from collections import defaultdict from copy import deepcopy from dataclasses import dataclass, field +from collections.abc import Hashable class CyclicException(Exception): @@ -15,6 +16,8 @@ class CyclicException(Exception): GraphType = defaultdict[str, set[str]] +EdgeType = tuple[str, str] +OwnerType = Hashable | None @dataclass @@ -28,6 +31,9 @@ class DAG: """ graph: GraphType = field(default_factory=lambda: defaultdict(set), init=False) + edge_owners: dict[EdgeType, set[OwnerType]] = field( + default_factory=dict, init=False + ) @property def nodes(self) -> set[str]: @@ -60,47 +66,94 @@ def invgraph(self) -> GraphType: invgraph[to_node].add(from_node) return invgraph - def add_edge(self, from_node: str, to_node: str) -> None: + def add_edge( + self, from_node: str, to_node: str, owner: OwnerType = None + ) -> bool: """ Ensure an edge exists in the graph. - Adds an edge from from_node to to_node. Does nothing if the edge already exists. + Adds an edge from from_node to to_node for the given owner. + If this is an additional owner for an existing edge, topology does not change. If the edge would make the graph cyclic, raises CyclicException. :param from_node: Source node name :type from_node: str :param to_node: Destination node name :type to_node: str + :param owner: Owner token for this edge; ``None`` is treated as persistent. + :type owner: collections.abc.Hashable | None :raises CyclicException: If adding the edge would create a cycle + :return: True if graph topology changed; False if this only added an owner. + :rtype: bool """ if from_node == to_node: raise CyclicException - test_graph = deepcopy(self.graph) - test_graph[from_node].add(to_node) - test_graph[to_node] + edge = (from_node, to_node) + owners = self.edge_owners.setdefault(edge, set()) + if owner in owners: + return False - if from_node in _bfs(test_graph, from_node): - raise CyclicException + topology_changed = len(owners) == 0 + if topology_changed: + test_graph = deepcopy(self.graph) + test_graph[from_node].add(to_node) + test_graph[to_node] + + if from_node in _bfs(test_graph, from_node): + if len(owners) == 0: + self.edge_owners.pop(edge, None) + raise CyclicException + + # No cycles! Modify referenced data structure + self.graph[from_node].add(to_node) + self.graph[to_node] - # No cycles! Modify referenced data structure - self.graph[from_node].add(to_node) - self.graph[to_node] + owners.add(owner) + return topology_changed - def remove_edge(self, from_node: str, to_node: str) -> None: + def remove_edge( + self, from_node: str, to_node: str, owner: OwnerType = None + ) -> bool: """ Ensure an edge is not present in the graph. - Removes an edge from from_node to to_node. Does nothing if the edge doesn't exist. + Removes ownership of an edge from from_node to to_node. + Topology only changes when the last owner is removed. Automatically prunes unconnected nodes after removal. :param from_node: Source node name :type from_node: str :param to_node: Destination node name :type to_node: str + :param owner: Owner token for this edge; ``None`` targets persistent ownership. + :type owner: collections.abc.Hashable | None + :return: True if graph topology changed; False if owner was absent or still shared. + :rtype: bool """ - self.graph.get(from_node, set()).discard(to_node) - self._prune() + edge = (from_node, to_node) + owners = self.edge_owners.get(edge, None) + if owners is None or owner not in owners: + return False + + owners.remove(owner) + + topology_changed = False + if len(owners) == 0: + self.edge_owners.pop(edge, None) + self.graph.get(from_node, set()).discard(to_node) + self._prune() + topology_changed = True + + return topology_changed + + def remove_owner(self, owner: OwnerType) -> set[EdgeType]: + removed_edges: set[EdgeType] = set() + for edge in list(self.edge_owners.keys()): + if owner in self.edge_owners.get(edge, set()): + if self.remove_edge(*edge, owner=owner): + removed_edges.add(edge) + return removed_edges def downstream(self, from_node: str) -> list[str]: """ diff --git a/src/ezmsg/core/graphcontext.py b/src/ezmsg/core/graphcontext.py index 1d62c2b2..81970387 100644 --- a/src/ezmsg/core/graphcontext.py +++ b/src/ezmsg/core/graphcontext.py @@ -1,38 +1,93 @@ import asyncio import logging import typing +import enum +import pickle -from .netprotocol import AddressType +from uuid import UUID +from types import TracebackType +from dataclasses import dataclass +from contextlib import suppress + +from .dag import CyclicException +from .netprotocol import ( + AddressType, + Command, + close_stream_writer, + encode_str, + read_int, + read_str, + uint64_to_bytes, +) from .graphserver import GraphServer, GraphService from .pubclient import Publisher from .subclient import Subscriber - -from types import TracebackType +from .graphmeta import ( + ProcessControlOperation, + GraphMetadata, + GraphSnapshot, + ProcessPing, + ProcessProfilingSnapshot, + ProcessProfilingTraceBatch, + ProfilingTraceStreamBatch, + ProfilingStreamControl, + ProcessStats, + ProcessControlResponse, + ProfilingTraceControl, + SettingsFieldUpdateRequest, + SettingsChangedEvent, + SettingsSnapshotValue, + TopologyChangedEvent, +) logger = logging.getLogger("ezmsg") -class GraphContext: - """ - GraphContext maintains a list of created publishers, subscribers, and connections in the graph. +class _SessionResponseKind(enum.Enum): + BYTE = enum.auto() + PICKLED = enum.auto() - The GraphContext provides a managed environment for creating and tracking publishers, - subscribers, and graph connections. When the context is no longer needed, it can - revert changes in the graph which disconnects publishers and removes modifications - that this context made. - It also maintains a context manager that ensures the GraphServer is running. +@dataclass +class _SessionCommand: + command: Command + args: tuple[str, ...] + payload: bytes | None + response_kind: _SessionResponseKind + response_fut: "asyncio.Future[typing.Any]" - :param graph_service: Optional graph service instance to use - :type graph_service: GraphService | None + +class GraphContext: + """ + Session-scoped client for graph mutation, metadata, settings, and process control. + + `GraphContext` opens a session connection to `GraphServer` and acts as a control + plane for both low-level graph operations and high-level API introspection. + + Core capabilities: + - Create/track `Publisher` and `Subscriber` clients. + - Connect/disconnect topic edges owned by this session. + - Register high-level `GraphMetadata`. + - Read graph snapshots (topology, edge ownership, sessions, process ownership). + - Query settings snapshots/events and subscribe to push-based settings updates. + - Route process-control requests (ping/stats/profiling and custom operations). + - Revert all session-owned mutations on context exit (`SESSION_CLEAR`). + + Session semantics: + - Mutations and metadata are tied to the session lifecycle. + - If the session disconnects, session-owned graph state is dropped by server cleanup. + - Low-level pub/sub API usage remains supported independently of metadata. + + :param graph_address: Graph server address. If `None`, defaults are used. + :type graph_address: AddressType | None :param auto_start: Whether to auto-start a GraphServer if connection fails. If None, defaults to auto-start only when graph_address is not provided and no environment override is set. :type auto_start: bool | None .. note:: - The GraphContext is typically managed automatically by the ezmsg runtime - and doesn't need to be instantiated directly by user code. + `GraphContext` is used by the runtime, and can also be used directly by tools + (inspectors, profilers, dashboards, and operational scripts). """ _clients: set[Publisher | Subscriber] @@ -40,6 +95,11 @@ class GraphContext: _graph_address: AddressType | None _graph_server: GraphServer | None + _session_id: UUID | None + _session_reader: asyncio.StreamReader | None + _session_writer: asyncio.StreamWriter | None + _session_task: asyncio.Task[None] | None + _session_commands: asyncio.Queue[_SessionCommand | None] | None def __init__( self, @@ -51,6 +111,11 @@ def __init__( self._graph_address = graph_address self._graph_server = None self._auto_start = auto_start + self._session_id = None + self._session_reader = None + self._session_writer = None + self._session_task = None + self._session_commands = None @property def graph_address(self) -> AddressType | None: @@ -98,8 +163,16 @@ async def connect(self, from_topic: str, to_topic: str) -> None: :param to_topic: The destination topic name :type to_topic: str """ - - await GraphService(self.graph_address).connect(from_topic, to_topic) + response = await self._session_command( + Command.SESSION_CONNECT, + from_topic, + to_topic, + response_kind=_SessionResponseKind.BYTE, + ) + if response == Command.CYCLIC.value: + raise CyclicException + if response != Command.COMPLETE.value: + raise RuntimeError("Unexpected response to session connect") self._edges.add((from_topic, to_topic)) async def disconnect(self, from_topic: str, to_topic: str) -> None: @@ -111,7 +184,14 @@ async def disconnect(self, from_topic: str, to_topic: str) -> None: :param to_topic: The destination topic name :type to_topic: str """ - await GraphService(self.graph_address).disconnect(from_topic, to_topic) + response = await self._session_command( + Command.SESSION_DISCONNECT, + from_topic, + to_topic, + response_kind=_SessionResponseKind.BYTE, + ) + if response != Command.COMPLETE.value: + raise RuntimeError("Unexpected response to session disconnect") self._edges.discard((from_topic, to_topic)) async def sync(self, timeout: float | None = None) -> None: @@ -140,6 +220,519 @@ async def _ensure_servers(self) -> None: auto_start=self._auto_start ) + async def _open_session(self) -> None: + if self._session_writer is not None: + return + + reader, writer = await GraphService(self.graph_address).open_connection() + writer.write(Command.SESSION.value) + await writer.drain() + + session_id = UUID(await read_str(reader)) + response = await reader.read(1) + if response != Command.COMPLETE.value: + await close_stream_writer(writer) + raise RuntimeError("Failed to create GraphContext session") + + self._session_id = session_id + self._session_reader = reader + self._session_writer = writer + self._session_commands = asyncio.Queue() + self._session_task = asyncio.create_task( + self._session_io_loop(), + name=f"graphctx-session-{session_id}", + ) + + def _require_session(self) -> tuple[asyncio.Queue[_SessionCommand | None], asyncio.Task[None]]: + if self._session_commands is None or self._session_task is None: + raise RuntimeError( + "GraphContext session is not active. Use GraphContext as an async context manager." + ) + return self._session_commands, self._session_task + + async def _session_io_loop(self) -> None: + reader = self._session_reader + writer = self._session_writer + commands = self._session_commands + if reader is None or writer is None or commands is None: + return + + try: + while True: + cmd = await commands.get() + if cmd is None: + break + + writer.write(cmd.command.value) + for arg in cmd.args: + writer.write(encode_str(arg)) + if cmd.payload is not None: + writer.write(uint64_to_bytes(len(cmd.payload))) + writer.write(cmd.payload) + await writer.drain() + + if cmd.response_kind == _SessionResponseKind.BYTE: + response = await reader.read(1) + + elif cmd.response_kind == _SessionResponseKind.PICKLED: + num_bytes = await read_int(reader) + payload_bytes = await reader.readexactly(num_bytes) + complete = await reader.read(1) + if complete != Command.COMPLETE.value: + raise RuntimeError("Unexpected pickled response from session") + response = pickle.loads(payload_bytes) + + else: + raise RuntimeError(f"Unsupported response kind: {cmd.response_kind}") + + if not cmd.response_fut.done(): + cmd.response_fut.set_result(response) + + except Exception as exc: + while True: + try: + pending = commands.get_nowait() + except asyncio.QueueEmpty: + break + + if pending is not None and not pending.response_fut.done(): + pending.response_fut.set_exception(exc) + finally: + while True: + try: + pending = commands.get_nowait() + except asyncio.QueueEmpty: + break + + if pending is not None and not pending.response_fut.done(): + pending.response_fut.set_exception( + RuntimeError("GraphContext session closed") + ) + + async def _close_session(self) -> None: + commands = self._session_commands + task = self._session_task + writer = self._session_writer + if writer is None: + return + + if commands is not None: + await commands.put(None) + if task is not None: + with suppress(asyncio.CancelledError): + await task + + await close_stream_writer(writer) + self._session_id = None + self._session_reader = None + self._session_writer = None + self._session_task = None + self._session_commands = None + self._edges.clear() + + async def _session_command( + self, + command: Command, + *args: str, + payload: bytes | None = None, + response_kind: _SessionResponseKind = _SessionResponseKind.BYTE, + ) -> typing.Any: + commands, task = self._require_session() + if task.done(): + raise RuntimeError("GraphContext session task is not running") + + response_fut: asyncio.Future[typing.Any] = asyncio.get_running_loop().create_future() + await commands.put( + _SessionCommand( + command=command, + args=tuple(args), + payload=payload, + response_kind=response_kind, + response_fut=response_fut, + ) + ) + return await response_fut + + async def register_metadata(self, metadata: GraphMetadata) -> None: + payload = pickle.dumps(metadata) + response = await self._session_command( + Command.SESSION_REGISTER, + payload=payload, + response_kind=_SessionResponseKind.BYTE, + ) + if response == Command.COMPLETE.value: + return + if response == Command.ERROR.value: + requested = set(metadata.components.keys()) + collisions: set[str] = set() + if len(requested) > 0: + own_session_id = str(self._session_id) if self._session_id is not None else None + try: + snapshot = await self.snapshot() + for session_id, session in snapshot.sessions.items(): + if own_session_id is not None and session_id == own_session_id: + continue + if session.metadata is None: + continue + collisions.update( + requested.intersection(session.metadata.components.keys()) + ) + except Exception: + # Fall back to a generic error if snapshot lookup fails. + pass + + if len(collisions) > 0: + collision_str = ", ".join(sorted(collisions)) + raise RuntimeError( + "Session metadata registration rejected by GraphServer due to " + f"component address collision(s): {collision_str}" + ) + raise RuntimeError("Session metadata registration rejected by GraphServer") + raise RuntimeError( + "Unexpected response to session metadata registration: " + f"{response!r}" + ) + + async def snapshot(self) -> GraphSnapshot: + snapshot = await self._session_command( + Command.SESSION_SNAPSHOT, + response_kind=_SessionResponseKind.PICKLED, + ) + if not isinstance(snapshot, GraphSnapshot): + raise RuntimeError("Session snapshot payload was not a GraphSnapshot") + return snapshot + + async def settings_snapshot(self) -> dict[str, SettingsSnapshotValue]: + snapshot = await self._session_command( + Command.SESSION_SETTINGS_SNAPSHOT, + response_kind=_SessionResponseKind.PICKLED, + ) + if not isinstance(snapshot, dict): + raise RuntimeError("Settings snapshot payload was not a dictionary") + if not all(isinstance(value, SettingsSnapshotValue) for value in snapshot.values()): + raise RuntimeError("Settings snapshot payload contained invalid values") + return snapshot + + async def settings_events(self, after_seq: int = 0) -> list[SettingsChangedEvent]: + events = await self._session_command( + Command.SESSION_SETTINGS_EVENTS, + str(after_seq), + response_kind=_SessionResponseKind.PICKLED, + ) + if not isinstance(events, list): + raise RuntimeError("Settings event payload was not a list") + if not all(isinstance(event, SettingsChangedEvent) for event in events): + raise RuntimeError("Settings event payload contained invalid entries") + return events + + async def settings_input_topic(self, component_address: str) -> str: + """ + Resolve the dynamic settings input topic for a component. + + The topic is discovered from currently registered session metadata. + Raises if the component is missing, does not opt in to dynamic settings, + or appears with conflicting dynamic settings topics. + """ + snapshot = await self.snapshot() + topics: set[str] = set() + for session in snapshot.sessions.values(): + metadata = session.metadata + if metadata is None: + continue + component = metadata.components.get(component_address) + if component is None: + continue + dynamic_settings = component.dynamic_settings + if dynamic_settings.enabled and dynamic_settings.input_topic is not None: + topics.add(dynamic_settings.input_topic) + + if len(topics) == 1: + return next(iter(topics)) + if len(topics) > 1: + raise RuntimeError( + "Conflicting dynamic settings topics for component " + f"'{component_address}': {sorted(topics)}" + ) + raise RuntimeError( + f"Component '{component_address}' does not expose dynamic settings metadata" + ) + + async def update_settings( + self, + component_address: str, + value: object, + *, + input_topic: str | None = None, + ) -> None: + """ + Publish a settings value to a component's `INPUT_SETTINGS` inlet. + + By default the target topic is resolved from metadata via + :meth:`settings_input_topic`. Supplying `input_topic` bypasses + metadata lookup. + """ + topic = input_topic if input_topic is not None else await self.settings_input_topic( + component_address + ) + pub = await self.publisher(topic) + try: + await pub.broadcast(value) + finally: + pub.close() + await pub.wait_closed() + self._clients.discard(pub) + + async def update_setting( + self, + component_address: str, + field_path: str, + value: object, + *, + timeout: float = 2.0, + ) -> SettingsSnapshotValue: + """ + Patch one field of a unit's current dynamic settings value. + + The patch is routed to the owning backend process, applied in-process + using dataclass replacement, and then published to `INPUT_SETTINGS`. + Returns a snapshot representation of the patched settings value. + """ + response = await self.process_request( + component_address, + ProcessControlOperation.UPDATE_SETTING_FIELD, + payload_obj=SettingsFieldUpdateRequest(field_path=field_path, value=value), + timeout=timeout, + ) + return typing.cast( + SettingsSnapshotValue, + self.decode_process_payload(response, SettingsSnapshotValue), + ) + + async def subscribe_settings_events( + self, + *, + after_seq: int = 0, + ) -> typing.AsyncIterator[SettingsChangedEvent]: + async for event in self._subscribe_pickled_stream( + command=Command.SESSION_SETTINGS_SUBSCRIBE, + setup_payload=encode_str(str(after_seq)), + expected_type=SettingsChangedEvent, + subscribe_error="Failed to subscribe to settings events", + payload_error="Settings subscription received invalid event payload", + ): + yield typing.cast(SettingsChangedEvent, event) + + async def subscribe_topology_events( + self, + *, + after_seq: int = 0, + ) -> typing.AsyncIterator[TopologyChangedEvent]: + async for event in self._subscribe_pickled_stream( + command=Command.SESSION_TOPOLOGY_SUBSCRIBE, + setup_payload=encode_str(str(after_seq)), + expected_type=TopologyChangedEvent, + subscribe_error="Failed to subscribe to topology events", + payload_error="Topology subscription received invalid event payload", + ): + yield typing.cast(TopologyChangedEvent, event) + + async def subscribe_profiling_trace( + self, + control: ProfilingStreamControl, + ) -> typing.AsyncIterator[ProfilingTraceStreamBatch]: + """ + Subscribe to streamed profiling trace batches from GraphServer. + """ + payload = pickle.dumps(control) + setup_payload = uint64_to_bytes(len(payload)) + payload + async for batch in self._subscribe_pickled_stream( + command=Command.SESSION_PROFILING_SUBSCRIBE, + setup_payload=setup_payload, + expected_type=ProfilingTraceStreamBatch, + subscribe_error="Failed to subscribe to profiling trace stream", + payload_error="Profiling subscription received invalid batch payload", + ): + yield typing.cast(ProfilingTraceStreamBatch, batch) + + async def _subscribe_pickled_stream( + self, + *, + command: Command, + setup_payload: bytes, + expected_type: type[object], + subscribe_error: str, + payload_error: str, + ) -> typing.AsyncIterator[object]: + reader, writer = await GraphService(self.graph_address).open_connection() + writer.write(command.value) + writer.write(setup_payload) + await writer.drain() + + _subscriber_id = UUID(await read_str(reader)) + response = await reader.read(1) + if response != Command.COMPLETE.value: + await close_stream_writer(writer) + raise RuntimeError(subscribe_error) + + try: + while True: + payload_size = await read_int(reader) + payload = await reader.readexactly(payload_size) + value = pickle.loads(payload) + if not isinstance(value, expected_type): + raise RuntimeError(payload_error) + yield value + except asyncio.IncompleteReadError: + return + finally: + await close_stream_writer(writer) + + async def process_request( + self, + unit_address: str, + operation: ProcessControlOperation | str, + *, + payload: bytes | None = None, + payload_obj: object | None = None, + timeout: float = 2.0, + ) -> ProcessControlResponse: + if payload is not None and payload_obj is not None: + raise ValueError("Specify only one of payload or payload_obj") + + if payload_obj is not None: + payload = pickle.dumps(payload_obj) + + operation_name = ( + operation.value if isinstance(operation, ProcessControlOperation) else operation + ) + response = await self._session_command( + Command.SESSION_PROCESS_REQUEST, + unit_address, + operation_name, + str(timeout), + payload=payload if payload is not None else b"", + response_kind=_SessionResponseKind.PICKLED, + ) + if not isinstance(response, ProcessControlResponse): + raise RuntimeError("Session process request payload was not ProcessControlResponse") + return response + + async def process_ping( + self, + unit_address: str, + *, + timeout: float = 2.0, + ) -> ProcessPing: + response = await self.process_request( + unit_address, + ProcessControlOperation.PING, + timeout=timeout, + ) + return typing.cast(ProcessPing, self.decode_process_payload(response, ProcessPing)) + + async def process_stats( + self, + unit_address: str, + *, + timeout: float = 2.0, + ) -> ProcessStats: + response = await self.process_request( + unit_address, + ProcessControlOperation.GET_PROCESS_STATS, + timeout=timeout, + ) + return typing.cast( + ProcessStats, self.decode_process_payload(response, ProcessStats) + ) + + async def process_profiling_snapshot( + self, + unit_address: str, + *, + timeout: float = 2.0, + ) -> ProcessProfilingSnapshot: + response = await self.process_request( + unit_address, + ProcessControlOperation.GET_PROFILING_SNAPSHOT, + timeout=timeout, + ) + return typing.cast( + ProcessProfilingSnapshot, + self.decode_process_payload(response, ProcessProfilingSnapshot), + ) + + async def process_set_profiling_trace( + self, + unit_address: str, + control: ProfilingTraceControl, + *, + timeout: float = 2.0, + ) -> ProcessControlResponse: + return await self.process_request( + unit_address, + ProcessControlOperation.SET_PROFILING_TRACE, + payload_obj=control, + timeout=timeout, + ) + + async def process_profiling_trace_batch( + self, + unit_address: str, + *, + max_samples: int = 1000, + timeout: float = 2.0, + ) -> ProcessProfilingTraceBatch: + response = await self.process_request( + unit_address, + ProcessControlOperation.GET_PROFILING_TRACE_BATCH, + payload_obj=max_samples, + timeout=timeout, + ) + return typing.cast( + ProcessProfilingTraceBatch, + self.decode_process_payload(response, ProcessProfilingTraceBatch), + ) + + async def profiling_snapshot_all( + self, + *, + timeout_per_process: float = 0.5, + ) -> dict[UUID, ProcessProfilingSnapshot]: + graph_snapshot = await self.snapshot() + out: dict[UUID, ProcessProfilingSnapshot] = {} + for process in graph_snapshot.processes.values(): + if len(process.units) == 0: + continue + route_unit = process.units[0] + try: + out[process.process_id] = await self.process_profiling_snapshot( + route_unit, timeout=timeout_per_process + ) + except Exception: + continue + return out + + def decode_process_payload( + self, + response: ProcessControlResponse, + expected_type: type[object] = object, + ) -> object: + if not response.ok: + raise RuntimeError( + f"Process request failed ({response.error_code}): {response.error}" + ) + if response.payload is None: + raise RuntimeError("Process response did not include a payload") + decoded = pickle.loads(response.payload) + if expected_type is object: + return decoded + if not isinstance(decoded, expected_type): + raise RuntimeError( + "Unexpected process payload type: " + f"{type(decoded).__name__} (expected {expected_type.__name__})" + ) + return decoded + async def _shutdown_servers(self) -> None: if self._graph_server is not None: self._graph_server.stop() @@ -147,6 +740,7 @@ async def _shutdown_servers(self) -> None: async def __aenter__(self) -> "GraphContext": await self._ensure_servers() + await self._open_session() return self async def __aexit__( @@ -156,6 +750,7 @@ async def __aexit__( exc_tb: TracebackType | None, ) -> bool: await self.revert() + await self._close_session() await self._shutdown_servers() return False @@ -174,8 +769,25 @@ async def revert(self) -> None: for future in asyncio.as_completed(wait): await future - for edge in self._edges: + self._clients.clear() + + if self._session_writer is not None: try: - await GraphService(self.graph_address).disconnect(*edge) - except (ConnectionRefusedError, BrokenPipeError, ConnectionResetError) as e: - logger.warn(f"Could not remove edge {edge} from GraphServer: {e}") + response = await self._session_command( + Command.SESSION_CLEAR, + response_kind=_SessionResponseKind.BYTE, + ) + if response != Command.COMPLETE.value: + logger.warning( + "GraphServer returned unexpected response to SESSION_CLEAR" + ) + except ( + ConnectionRefusedError, + BrokenPipeError, + ConnectionResetError, + asyncio.IncompleteReadError, + RuntimeError, + ) as e: + logger.warning(f"Could not clear GraphContext session state: {e}") + + self._edges.clear() diff --git a/src/ezmsg/core/graphmeta.py b/src/ezmsg/core/graphmeta.py new file mode 100644 index 00000000..7bb0973f --- /dev/null +++ b/src/ezmsg/core/graphmeta.py @@ -0,0 +1,379 @@ +import enum + +from dataclasses import dataclass, field +from typing import Any, TypeAlias, NamedTuple +from uuid import UUID + + +@dataclass +class DynamicSettingsMetadata: + enabled: bool + input_topic: str | None + settings_type: str | None + + +@dataclass +class StreamMetadata: + name: str + address: str + msg_type: str + + +@dataclass +class InputStreamMetadata(StreamMetadata): + leaky: bool = False + max_queue: int | None = None + + +@dataclass +class OutputStreamMetadata(StreamMetadata): + host: str | None = None + port: int | None = None + num_buffers: int | None = None + buf_size: int | None = None + force_tcp: bool | None = None + + +StreamMetadataType: TypeAlias = ( + StreamMetadata | InputStreamMetadata | OutputStreamMetadata +) + + +@dataclass +class TopicMetadata: + name: str + address: str + msg_type: str + + +@dataclass +class InputTopicMetadata(TopicMetadata): ... + + +@dataclass +class OutputTopicMetadata(TopicMetadata): ... + + +TopicMetadataType: TypeAlias = TopicMetadata | InputTopicMetadata | OutputTopicMetadata + + +@dataclass +class InputRelayMetadata(InputTopicMetadata): + leaky: bool = False + max_queue: int | None = None + copy_on_forward: bool = True + + +@dataclass +class OutputRelayMetadata(OutputTopicMetadata): + host: str | None = None + port: int | None = None + num_buffers: int | None = None + buf_size: int | None = None + force_tcp: bool | None = None + copy_on_forward: bool = True + + +RelayMetadataType: TypeAlias = InputRelayMetadata | OutputRelayMetadata + + +@dataclass +class TaskMetadata: + name: str + subscribes: str | None = None + publishes: list[str] = field(default_factory=list) + + +@dataclass +class SettingsFieldMetadata: + name: str + field_type: str + required: bool + default: Any + description: str | None + bounds: tuple[float | None, float | None] | None + choices: list[Any] | None + widget_hint: str | None + + +@dataclass +class SettingsSchemaMetadata: + provider: str + settings_type: str + fields: list[SettingsFieldMetadata] + + +SettingsReprType: TypeAlias = dict[str, Any] | str +SerializedSettingsType: TypeAlias = bytes | None +InitialSettingsType: TypeAlias = tuple[SerializedSettingsType, SettingsReprType] + + +@dataclass +class ComponentMetadata: + address: str + name: str + component_type: str + settings_type: str + initial_settings: InitialSettingsType + dynamic_settings: DynamicSettingsMetadata + settings_schema: SettingsSchemaMetadata | None + + +@dataclass +class CollectionMetadata(ComponentMetadata): + topics: dict[str, TopicMetadataType] + relays: dict[str, RelayMetadataType] + children: list[str] + + +@dataclass +class UnitMetadata(ComponentMetadata): + streams: dict[str, StreamMetadataType] + tasks: list[TaskMetadata] + main: str | None + threads: list[str] + + +ComponentMetadataType: TypeAlias = ( + ComponentMetadata | CollectionMetadata | UnitMetadata +) + + +@dataclass +class GraphMetadata: + schema_version: int + root_name: str | None + components: dict[str, ComponentMetadataType] + + +@dataclass +class ProcessRegistration: + pid: int + host: str + units: list[str] + + +@dataclass +class ProcessOwnershipUpdate: + added_units: list[str] = field(default_factory=list) + removed_units: list[str] = field(default_factory=list) + + +@dataclass +class SettingsSnapshotValue: + serialized: bytes | None + repr_value: dict[str, Any] | str + structured_value: dict[str, Any] | None = None + settings_schema: SettingsSchemaMetadata | None = None + + +class SettingsEventType(enum.Enum): + INITIAL_SETTINGS = "INITIAL_SETTINGS" + SETTINGS_UPDATED = "SETTINGS_UPDATED" + + +@dataclass +class SettingsChangedEvent: + seq: int + event_type: SettingsEventType + component_address: str + timestamp: float + source_session_id: str | None + source_process_id: UUID | None + value: SettingsSnapshotValue + + +class TopologyEventType(enum.Enum): + GRAPH_CHANGED = "GRAPH_CHANGED" + PROCESS_CHANGED = "PROCESS_CHANGED" + + +@dataclass +class TopologyChangedEvent: + seq: int + event_type: TopologyEventType + timestamp: float + changed_topics: list[str] + source_session_id: str | None + source_process_id: UUID | None + + +@dataclass +class ProcessSettingsUpdate: + component_address: str + value: SettingsSnapshotValue + timestamp: float + + +@dataclass +class ProcessControlRequest: + request_id: str + unit_address: str + operation: "ProcessControlOperation | str" + payload: bytes | None = None + + +class ProcessControlOperation(enum.Enum): + PING = "PING" + GET_PROCESS_STATS = "GET_PROCESS_STATS" + GET_PROFILING_SNAPSHOT = "GET_PROFILING_SNAPSHOT" + SET_PROFILING_TRACE = "SET_PROFILING_TRACE" + GET_PROFILING_TRACE_BATCH = "GET_PROFILING_TRACE_BATCH" + UPDATE_SETTING_FIELD = "UPDATE_SETTING_FIELD" + + +class ProcessControlErrorCode(enum.Enum): + UNROUTABLE_UNIT = "UNROUTABLE_UNIT" + ROUTE_WRITE_FAILED = "ROUTE_WRITE_FAILED" + TIMEOUT = "TIMEOUT" + PROCESS_DISCONNECTED = "PROCESS_DISCONNECTED" + UNSUPPORTED_OPERATION = "UNSUPPORTED_OPERATION" + HANDLER_NOT_CONFIGURED = "HANDLER_NOT_CONFIGURED" + HANDLER_ERROR = "HANDLER_ERROR" + INVALID_RESPONSE = "INVALID_RESPONSE" + + +@dataclass +class ProcessControlResponse: + request_id: str + ok: bool + payload: bytes | None = None + error: str | None = None + error_code: ProcessControlErrorCode | None = None + process_id: UUID | None = None + + +@dataclass +class SettingsFieldUpdateRequest: + field_path: str + value: Any + + +@dataclass +class ProcessPing: + process_id: UUID + pid: int + host: str + timestamp: float + + +@dataclass +class ProcessStats: + process_id: UUID + pid: int + host: str + owned_units: list[str] + timestamp: float + + +class ProfileChannelType(enum.Enum): + LOCAL = "LOCAL" + SHM = "SHM" + TCP = "TCP" + UNKNOWN = "UNKNOWN" + + +@dataclass +class PublisherProfileSnapshot: + endpoint_id: str + topic: str + messages_published_total: int + messages_published_window: int + publish_rate_hz_window: float + inflight_messages_current: int + num_buffers: int + timestamp: float + + +@dataclass +class SubscriberProfileSnapshot: + endpoint_id: str + topic: str + messages_received_total: int + messages_received_window: int + channel_kind_last: ProfileChannelType + timestamp: float + + +@dataclass +class ProcessProfilingSnapshot: + process_id: UUID + pid: int + host: str + window_seconds: float + timestamp: float + publishers: dict[str, PublisherProfileSnapshot] + subscribers: dict[str, SubscriberProfileSnapshot] + + +@dataclass +class ProfilingTraceControl: + enabled: bool + sample_mod: int = 1 + publisher_topics: list[str] | None = None + subscriber_topics: list[str] | None = None + publisher_endpoint_ids: list[str] | None = None + subscriber_endpoint_ids: list[str] | None = None + metrics: list[str] | None = None + ttl_seconds: float | None = None + + +@dataclass +class ProfilingTraceSample: + timestamp: float + endpoint_id: str + topic: str + metric: str + value: float + channel_kind: ProfileChannelType | None = None + sample_seq: int | None = None + + +@dataclass +class ProcessProfilingTraceBatch: + process_id: UUID + pid: int + host: str + timestamp: float + samples: list[ProfilingTraceSample] + + +@dataclass +class ProfilingTraceStreamBatch: + timestamp: float + batches: dict[UUID, ProcessProfilingTraceBatch] + + +@dataclass +class ProfilingStreamControl: + interval: float = 0.05 + max_samples: int = 1000 + process_ids: list[UUID] | None = None + include_empty_batches: bool = False + + +class Edge(NamedTuple): + from_topic: str + to_topic: str + + +@dataclass +class SnapshotSession: + edges: list[Edge] + metadata: GraphMetadata | None + + +@dataclass +class SnapshotProcess: + process_id: UUID + pid: int | None + host: str | None + units: list[str] + + +@dataclass +class GraphSnapshot: + graph: dict[str, list[str]] + edge_owners: dict[Edge, list[str]] + sessions: dict[str, SnapshotSession] + processes: dict[UUID, SnapshotProcess] = field(default_factory=dict) diff --git a/src/ezmsg/core/graphserver.py b/src/ezmsg/core/graphserver.py index f4abf8fc..aa04d2b7 100644 --- a/src/ezmsg/core/graphserver.py +++ b/src/ezmsg/core/graphserver.py @@ -4,17 +4,46 @@ import os import socket import threading +import time +from collections import deque +from collections.abc import Sequence from contextlib import suppress from uuid import UUID, uuid1 +from dataclasses import dataclass from . import __version__ from .dag import DAG, CyclicException from .graph_util import get_compactified_graph, graph_string, prune_graph_connections +from .graphmeta import ( + Edge, + ProcessControlOperation, + ProcessControlErrorCode, + GraphMetadata, + GraphSnapshot, + ProcessProfilingTraceBatch, + ProfilingTraceSample, + ProfilingTraceStreamBatch, + ProfilingStreamControl, + ProcessControlRequest, + ProcessControlResponse, + ProcessRegistration, + ProcessOwnershipUpdate, + ProcessSettingsUpdate, + SettingsChangedEvent, + SettingsEventType, + SettingsSnapshotValue, + TopologyChangedEvent, + TopologyEventType, + SnapshotProcess, + SnapshotSession, +) from .netprotocol import ( Address, Command, ClientInfo, + ProcessInfo, + SessionInfo, SubscriberInfo, PublisherInfo, ChannelInfo, @@ -35,6 +64,30 @@ from .shm import SHMContext, SHMInfo logger = logging.getLogger("ezmsg") +PERSISTENT_EDGE_OWNER = None +SUBSCRIBER_UPDATE_TIMEOUT_SEC = float( + os.environ.get("EZMSG_SUBSCRIBER_UPDATE_TIMEOUT_SEC", "1.0") +) + + +@dataclass +class _SettingsState: + value: SettingsSnapshotValue + metadata_session_id: UUID | None + source_process_id: UUID | None + + +@dataclass(frozen=True) +class _RetentionPolicy: + profiling_trace_buffer_limit: int = int( + os.environ.get("EZMSG_PROFILE_TRACE_BUFFER_LIMIT", "200000") + ) + settings_event_history_limit: int = int( + os.environ.get("EZMSG_SETTINGS_EVENT_HISTORY_LIMIT", "10000") + ) + topology_event_history_limit: int = int( + os.environ.get("EZMSG_TOPOLOGY_EVENT_HISTORY_LIMIT", "10000") + ) class GraphServer(threading.Thread): @@ -65,6 +118,21 @@ class GraphServer(threading.Thread): _client_tasks: dict[UUID, "asyncio.Task[None]"] _command_lock: asyncio.Lock + _settings_state: dict[str, _SettingsState] + _settings_events: list[SettingsChangedEvent] + _settings_event_seq: int + _settings_owned_by_session: dict[UUID, set[str]] + _settings_subscribers: dict[UUID, asyncio.Queue[object]] + _topology_events: list[TopologyChangedEvent] + _topology_event_seq: int + _topology_subscribers: dict[UUID, asyncio.Queue[object]] + _pending_process_requests: dict[ + str, tuple[UUID, "asyncio.Future[ProcessControlResponse]"] + ] + _profiling_trace_buffers: dict[UUID, deque[tuple[int, ProfilingTraceSample]]] + _profiling_trace_process_meta: dict[UUID, tuple[int, str]] + _profiling_trace_seq: dict[UUID, int] + _retention_policy: _RetentionPolicy def __init__(self, **kwargs) -> None: super().__init__( @@ -80,6 +148,19 @@ def __init__(self, **kwargs) -> None: self._client_tasks = {} self.shms = {} self._address = None + self._settings_state = {} + self._settings_events = [] + self._settings_event_seq = 0 + self._settings_owned_by_session = {} + self._settings_subscribers = {} + self._topology_events = [] + self._topology_event_seq = 0 + self._topology_subscribers = {} + self._pending_process_requests = {} + self._profiling_trace_buffers = {} + self._profiling_trace_process_meta = {} + self._profiling_trace_seq = {} + self._retention_policy = _RetentionPolicy() @property def address(self) -> Address: @@ -261,6 +342,85 @@ async def api( # to avoid closing writer return + elif req == Command.SESSION.value: + session_id = uuid1() + self.clients[session_id] = SessionInfo(session_id, writer) + writer.write(encode_str(str(session_id))) + writer.write(Command.COMPLETE.value) + await writer.drain() + self._client_tasks[session_id] = asyncio.create_task( + self._handle_session(session_id, reader, writer) + ) + + # NOTE: Created a session client, must return early + # to avoid closing writer + return + + elif req == Command.SESSION_SETTINGS_SUBSCRIBE.value: + subscriber_id = uuid1() + after_seq = int(await read_str(reader)) + writer.write(encode_str(str(subscriber_id))) + writer.write(Command.COMPLETE.value) + await writer.drain() + self._client_tasks[subscriber_id] = asyncio.create_task( + self._handle_settings_subscriber( + subscriber_id, after_seq, reader, writer + ) + ) + + # NOTE: Created a stream client, must return early + # to avoid closing writer + return + + elif req == Command.SESSION_TOPOLOGY_SUBSCRIBE.value: + subscriber_id = uuid1() + after_seq = int(await read_str(reader)) + writer.write(encode_str(str(subscriber_id))) + writer.write(Command.COMPLETE.value) + await writer.drain() + self._client_tasks[subscriber_id] = asyncio.create_task( + self._handle_topology_subscriber( + subscriber_id, after_seq, reader, writer + ) + ) + + # NOTE: Created a stream client, must return early + # to avoid closing writer + return + + elif req == Command.SESSION_PROFILING_SUBSCRIBE.value: + subscriber_id = uuid1() + stream_control = await self._read_profiling_stream_control(reader) + writer.write(encode_str(str(subscriber_id))) + writer.write(Command.COMPLETE.value) + await writer.drain() + self._client_tasks[subscriber_id] = asyncio.create_task( + self._handle_profiling_subscriber( + subscriber_id, + stream_control, + reader, + writer, + ) + ) + + # NOTE: Created a stream client, must return early + # to avoid closing writer + return + + elif req == Command.PROCESS.value: + process_client_id = uuid1() + self.clients[process_client_id] = ProcessInfo(process_client_id, writer) + writer.write(encode_str(str(process_client_id))) + writer.write(Command.COMPLETE.value) + await writer.drain() + self._client_tasks[process_client_id] = asyncio.create_task( + self._handle_process(process_client_id, reader, writer) + ) + + # NOTE: Created a process control client, must return early + # to avoid closing writer + return + else: # We only want to handle one command at a time async with self._command_lock: @@ -302,23 +462,31 @@ async def api( elif req in [Command.CONNECT.value, Command.DISCONNECT.value]: from_topic = await read_str(reader) to_topic = await read_str(reader) - - cmd = self.graph.add_edge - if req == Command.DISCONNECT.value: - cmd = self.graph.remove_edge + topology_changed = False try: - cmd(from_topic, to_topic) - for sub in self._downstream_subs(to_topic): - await self._notify_subscriber(sub) + if req == Command.CONNECT.value: + topology_changed = self._connect_owner( + from_topic, to_topic, PERSISTENT_EDGE_OWNER + ) + else: + topology_changed = self._disconnect_owner( + from_topic, to_topic, PERSISTENT_EDGE_OWNER + ) writer.write(Command.COMPLETE.value) except CyclicException: writer.write(Command.CYCLIC.value) - await writer.drain() + if topology_changed: + self._append_topology_event_locked( + event_type=TopologyEventType.GRAPH_CHANGED, + changed_topics=[to_topic], + source_session_id=None, + source_process_id=None, + ) + await self._notify_downstream_for_topic(to_topic) - if req == Command.DISCONNECT.value: - await close_stream_writer(writer) + await writer.drain() elif req == Command.SYNC.value: for pub in self._publishers(): @@ -393,24 +561,1201 @@ async def _handle_client( finally: # Ensure any waiter on this client unblocks - # with suppress(Exception): self.clients[client_id].set_sync() self.clients.pop(client_id, None) await close_stream_writer(writer) + async def _handle_session( + self, + session_id: UUID, + reader: asyncio.StreamReader, + writer: asyncio.StreamWriter, + ) -> None: + logger.debug(f"Graph Server: Session connected: {session_id}") + + try: + while True: + req = await reader.read(1) + + if not req: + break + + if req in [ + Command.SESSION_CONNECT.value, + Command.SESSION_DISCONNECT.value, + ]: + response = await self._handle_session_edge_request( + session_id, req, reader + ) + writer.write(response) + await writer.drain() + + elif req == Command.SESSION_CLEAR.value: + response = await self._handle_session_clear_request(session_id) + writer.write(response) + await writer.drain() + + elif req == Command.SESSION_REGISTER.value: + response = await self._handle_session_register_request( + session_id, reader + ) + writer.write(response) + await writer.drain() + + elif req == Command.SESSION_SNAPSHOT.value: + await self._handle_session_snapshot_request(writer) + await writer.drain() + + elif req == Command.SESSION_SETTINGS_SNAPSHOT.value: + await self._handle_session_settings_snapshot_request(writer) + await writer.drain() + + elif req == Command.SESSION_SETTINGS_EVENTS.value: + after_seq = int(await read_str(reader)) + await self._handle_session_settings_events_request( + writer, after_seq + ) + await writer.drain() + + elif req == Command.SESSION_PROCESS_REQUEST.value: + await self._handle_session_process_request(writer, reader) + await writer.drain() + + else: + logger.warning( + f"Session {session_id} rx unknown command from GraphServer: {req}" + ) + + except (ConnectionResetError, BrokenPipeError) as e: + logger.debug(f"Session {session_id} disconnected from GraphServer: {e}") + + finally: + async with self._command_lock: + notify_topics = self._drop_session(session_id) + + for topic in notify_topics: + await self._notify_downstream_for_topic(topic) + + self._client_tasks.pop(session_id, None) + await close_stream_writer(writer) + + def _process_info(self, process_client_id: UUID) -> ProcessInfo | None: + info = self.clients.get(process_client_id) + if isinstance(info, ProcessInfo): + return info + return None + + def _process_key(self, process_client_id: UUID) -> UUID: + return process_client_id + + async def _handle_process( + self, + process_client_id: UUID, + reader: asyncio.StreamReader, + writer: asyncio.StreamWriter, + ) -> None: + logger.debug(f"Graph Server: Process control connected: {process_client_id}") + + try: + while True: + req = await reader.read(1) + + if not req: + break + + if req == Command.PROCESS_REGISTER.value: + response = await self._handle_process_register_request( + process_client_id, reader + ) + await self._write_process_response( + process_client_id, writer, response + ) + + elif req == Command.PROCESS_UPDATE_OWNERSHIP.value: + response = await self._handle_process_update_ownership_request( + process_client_id, reader + ) + await self._write_process_response( + process_client_id, writer, response + ) + + elif req == Command.PROCESS_SETTINGS_UPDATE.value: + response = await self._handle_process_settings_update_request( + process_client_id, reader + ) + await self._write_process_response( + process_client_id, writer, response + ) + + elif req == Command.PROCESS_PROFILING_TRACE_UPDATE.value: + await self._handle_process_profiling_trace_update_request( + process_client_id, reader + ) + + elif req == Command.PROCESS_ROUTE_RESPONSE.value: + await self._handle_process_route_response_request( + process_client_id, reader + ) + + else: + logger.warning( + f"Process control {process_client_id} rx unknown command: {req}" + ) + + except (ConnectionResetError, BrokenPipeError) as e: + logger.debug( + f"Process control {process_client_id} disconnected from GraphServer: {e}" + ) + + finally: + process_info = self._process_info(process_client_id) + + async with self._command_lock: + request_ids = [ + request_id + for request_id, (owner_process_id, _) in self._pending_process_requests.items() + if owner_process_id == process_client_id + ] + for request_id in request_ids: + pending = self._pending_process_requests.pop(request_id, None) + if pending is None: + continue + _, response_fut = pending + if not response_fut.done(): + response_fut.set_result( + ProcessControlResponse( + request_id=request_id, + ok=False, + error="Owning process disconnected before response", + error_code=ProcessControlErrorCode.PROCESS_DISCONNECTED, + process_id=self._process_key(process_client_id), + ) + ) + self._remove_settings_for_process_locked(process_client_id) + if process_info is not None: + source_process_id = self._process_key(process_client_id) + self._profiling_trace_buffers.pop(source_process_id, None) + self._profiling_trace_process_meta.pop(source_process_id, None) + self._profiling_trace_seq.pop(source_process_id, None) + self._append_topology_event_locked( + event_type=TopologyEventType.PROCESS_CHANGED, + changed_topics=[], + source_session_id=None, + source_process_id=source_process_id, + ) + self.clients.pop(process_client_id, None) + self._client_tasks.pop(process_client_id, None) + await close_stream_writer(writer) + + async def _write_process_response( + self, + process_client_id: UUID, + fallback_writer: asyncio.StreamWriter, + response: bytes, + ) -> None: + process_info = self._process_info(process_client_id) + if process_info is None: + fallback_writer.write(response) + await fallback_writer.drain() + return + + async with process_info.write_lock: + writer = process_info.writer + writer.write(response) + await writer.drain() + + async def _read_pickled_payload(self, reader: asyncio.StreamReader) -> object: + payload_size = await read_int(reader) + payload = await reader.readexactly(payload_size) + return pickle.loads(payload) + + async def _read_typed_payload( + self, + reader: asyncio.StreamReader, + expected_type: type[object], + *, + log_prefix: str, + ) -> object | None: + try: + payload_obj = await self._read_pickled_payload(reader) + if not isinstance(payload_obj, expected_type): + raise RuntimeError( + f"payload was not {expected_type.__name__}: {type(payload_obj).__name__}" + ) + return payload_obj + except Exception as exc: + logger.warning("%s parse failed; ignoring payload: %s", log_prefix, exc) + return None + + def _queue_stream_event( + self, + queue: asyncio.Queue[object], + event: object, + ) -> None: + try: + queue.put_nowait(event) + except asyncio.QueueFull: + # Keep most recent samples under backpressure. + with suppress(asyncio.QueueEmpty): + queue.get_nowait() + with suppress(asyncio.QueueFull): + queue.put_nowait(event) + + async def _stream_sender( + self, + subscriber_id: UUID, + queue: asyncio.Queue[object], + writer: asyncio.StreamWriter, + label: str, + ) -> None: + try: + while True: + event = await queue.get() + payload = pickle.dumps(event) + writer.write(uint64_to_bytes(len(payload))) + writer.write(payload) + await writer.drain() + except (ConnectionResetError, BrokenPipeError): + logger.debug(f"{label} subscriber {subscriber_id} disconnected on send") + except asyncio.CancelledError: + raise + + async def _handle_event_subscriber( + self, + *, + subscriber_id: UUID, + after_seq: int, + reader: asyncio.StreamReader, + writer: asyncio.StreamWriter, + queue: asyncio.Queue[object], + subscribers: dict[UUID, asyncio.Queue[object]], + events: Sequence[object], + label: str, + ) -> None: + async with self._command_lock: + subscribers[subscriber_id] = queue + for event in events: + if getattr(event, "seq", 0) > after_seq: + self._queue_stream_event(queue, event) + + sender_task = asyncio.create_task( + self._stream_sender(subscriber_id, queue, writer, label), + name=f"{label}-sender-{subscriber_id}", + ) + + try: + while True: + req = await reader.read(1) + if not req: + break + except (ConnectionResetError, BrokenPipeError) as e: + logger.debug(f"{label} subscriber {subscriber_id} disconnected: {e}") + finally: + async with self._command_lock: + subscribers.pop(subscriber_id, None) + self._client_tasks.pop(subscriber_id, None) + sender_task.cancel() + with suppress(asyncio.CancelledError): + await sender_task + await close_stream_writer(writer) + + async def _handle_settings_subscriber( + self, + subscriber_id: UUID, + after_seq: int, + reader: asyncio.StreamReader, + writer: asyncio.StreamWriter, + ) -> None: + queue: asyncio.Queue[object] = asyncio.Queue(maxsize=1024) + await self._handle_event_subscriber( + subscriber_id=subscriber_id, + after_seq=after_seq, + reader=reader, + writer=writer, + queue=queue, + subscribers=self._settings_subscribers, + events=self._settings_events, + label="settings", + ) + + async def _handle_topology_subscriber( + self, + subscriber_id: UUID, + after_seq: int, + reader: asyncio.StreamReader, + writer: asyncio.StreamWriter, + ) -> None: + queue: asyncio.Queue[object] = asyncio.Queue(maxsize=1024) + await self._handle_event_subscriber( + subscriber_id=subscriber_id, + after_seq=after_seq, + reader=reader, + writer=writer, + queue=queue, + subscribers=self._topology_subscribers, + events=self._topology_events, + label="topology", + ) + + async def _read_profiling_stream_control( + self, reader: asyncio.StreamReader + ) -> ProfilingStreamControl: + payload_obj = await self._read_pickled_payload(reader) + if not isinstance(payload_obj, ProfilingStreamControl): + raise RuntimeError( + "Invalid profiling stream control payload type: " + f"{type(payload_obj).__name__}" + ) + return payload_obj + + async def _collect_profiling_trace_stream_batch( + self, + *, + stream_control: ProfilingStreamControl, + last_seq_by_process: dict[UUID, int], + ) -> ProfilingTraceStreamBatch: + process_ids_filter = ( + set(stream_control.process_ids) + if stream_control.process_ids is not None + else None + ) + max_samples = max(1, int(stream_control.max_samples)) + now_ts = time.time() + batches: dict[UUID, ProcessProfilingTraceBatch] = {} + + async with self._command_lock: + connected_processes: dict[UUID, tuple[int, str]] = {} + for client_id, info in self.clients.items(): + if not isinstance(info, ProcessInfo): + continue + process_id = self._process_key(client_id) + pid = info.pid if info.pid is not None else -1 + host = info.host if info.host is not None else "" + connected_processes[process_id] = (pid, host) + + process_ids: list[UUID] + if process_ids_filter is not None: + process_ids = sorted(process_ids_filter, key=str) + else: + process_ids = sorted(connected_processes.keys(), key=str) + + for process_id in process_ids: + sample_buffer = self._profiling_trace_buffers.get(process_id) + samples: list[ProfilingTraceSample] = [] + if sample_buffer: + last_seq = last_seq_by_process.get(process_id, 0) + oldest_seq = sample_buffer[0][0] + if last_seq < oldest_seq - 1: + last_seq = oldest_seq - 1 + for seq, sample in sample_buffer: + if seq <= last_seq: + continue + samples.append(sample) + last_seq = seq + if len(samples) >= max_samples: + break + last_seq_by_process[process_id] = last_seq + + if len(samples) == 0 and not stream_control.include_empty_batches: + continue + + pid, host = connected_processes.get( + process_id, + self._profiling_trace_process_meta.get(process_id, (-1, "")), + ) + batches[process_id] = ProcessProfilingTraceBatch( + process_id=process_id, + pid=pid, + host=host, + timestamp=now_ts, + samples=samples, + ) + + return ProfilingTraceStreamBatch(timestamp=now_ts, batches=batches) + + async def _handle_profiling_subscriber( + self, + subscriber_id: UUID, + stream_control: ProfilingStreamControl, + reader: asyncio.StreamReader, + writer: asyncio.StreamWriter, + ) -> None: + interval = max(0.01, float(stream_control.interval)) + last_seq_by_process: dict[UUID, int] = {} + try: + while True: + try: + req = await asyncio.wait_for(reader.read(1), timeout=interval) + if not req: + break + # No control commands currently supported on this stream. + continue + except asyncio.TimeoutError: + pass + + batch = await self._collect_profiling_trace_stream_batch( + stream_control=stream_control, + last_seq_by_process=last_seq_by_process, + ) + if len(batch.batches) == 0: + continue + + payload = pickle.dumps(batch) + writer.write(uint64_to_bytes(len(payload))) + writer.write(payload) + await writer.drain() + except (ConnectionResetError, BrokenPipeError) as e: + logger.debug(f"Profiling subscriber {subscriber_id} disconnected: {e}") + except asyncio.CancelledError: + raise + except Exception as exc: + logger.error( + "Profiling subscriber %s failed: %s", + subscriber_id, + exc, + ) + finally: + self._client_tasks.pop(subscriber_id, None) + await close_stream_writer(writer) + + async def _handle_process_profiling_trace_update_request( + self, process_client_id: UUID, reader: asyncio.StreamReader + ) -> None: + batch = await self._read_typed_payload( + reader, + ProcessProfilingTraceBatch, + log_prefix=f"Process control {process_client_id} trace update", + ) + if batch is None: + return + + async with self._command_lock: + process_id = self._process_key(process_client_id) + trace_buffer = self._profiling_trace_buffers.setdefault( + process_id, + deque(maxlen=self._retention_policy.profiling_trace_buffer_limit), + ) + next_seq = self._profiling_trace_seq.get(process_id, 0) + for sample in batch.samples: + next_seq += 1 + trace_buffer.append((next_seq, sample)) + self._profiling_trace_seq[process_id] = next_seq + self._profiling_trace_process_meta[process_id] = (batch.pid, batch.host) + + async def _handle_process_register_request( + self, process_client_id: UUID, reader: asyncio.StreamReader + ) -> bytes: + registration = await self._read_typed_payload( + reader, + ProcessRegistration, + log_prefix=f"Process control {process_client_id} registration", + ) + if registration is None: + return Command.ERROR.value + + async with self._command_lock: + process_info = self._process_info(process_client_id) + if process_info is None: + return Command.ERROR.value + + conflicts = sorted( + { + unit + for unit in set(registration.units) + if ( + (owner := self._process_owner_for_unit(unit)) is not None + and owner != process_client_id + ) + } + ) + if conflicts: + logger.warning( + "Process control %s register rejected due to unit ownership conflict(s): %s", + process_client_id, + ", ".join(conflicts), + ) + return Command.ERROR.value + + prev_units = set(process_info.units) + process_info.pid = registration.pid + process_info.host = registration.host + process_info.units = set(registration.units) + if prev_units != process_info.units: + self._append_topology_event_locked( + event_type=TopologyEventType.PROCESS_CHANGED, + changed_topics=[], + source_session_id=None, + source_process_id=self._process_key(process_client_id), + ) + + return Command.COMPLETE.value + + async def _handle_process_update_ownership_request( + self, process_client_id: UUID, reader: asyncio.StreamReader + ) -> bytes: + update = await self._read_typed_payload( + reader, + ProcessOwnershipUpdate, + log_prefix=f"Process control {process_client_id} ownership update", + ) + if update is None: + return Command.ERROR.value + + async with self._command_lock: + process_info = self._process_info(process_client_id) + if process_info is None: + return Command.ERROR.value + + conflicts = sorted( + { + unit + for unit in set(update.added_units) + if ( + (owner := self._process_owner_for_unit(unit)) is not None + and owner != process_client_id + ) + } + ) + if conflicts: + logger.warning( + "Process control %s ownership update rejected due to unit ownership conflict(s): %s", + process_client_id, + ", ".join(conflicts), + ) + return Command.ERROR.value + + prev_units = set(process_info.units) + process_info.units.update(update.added_units) + process_info.units.difference_update(update.removed_units) + if prev_units != process_info.units: + self._append_topology_event_locked( + event_type=TopologyEventType.PROCESS_CHANGED, + changed_topics=[], + source_session_id=None, + source_process_id=self._process_key(process_client_id), + ) + + return Command.COMPLETE.value + + async def _handle_process_settings_update_request( + self, process_client_id: UUID, reader: asyncio.StreamReader + ) -> bytes: + update = await self._read_typed_payload( + reader, + ProcessSettingsUpdate, + log_prefix=f"Process control {process_client_id} settings update", + ) + if update is None: + return Command.ERROR.value + + async with self._command_lock: + process_info = self._process_info(process_client_id) + if process_info is None: + return Command.ERROR.value + if update.component_address not in process_info.units: + logger.warning( + "Process control %s settings update rejected for unowned component: %s", + process_client_id, + update.component_address, + ) + return Command.ERROR.value + + prior_state = self._settings_state.get(update.component_address) + metadata_session_id = self._session_owner_for_component_locked( + update.component_address + ) + if metadata_session_id is None and prior_state is not None: + metadata_session_id = prior_state.metadata_session_id + + source_process_id = self._process_key(process_client_id) + self._settings_state[update.component_address] = _SettingsState( + value=update.value, + metadata_session_id=metadata_session_id, + source_process_id=source_process_id, + ) + self._append_settings_event_locked( + event_type=SettingsEventType.SETTINGS_UPDATED, + component_address=update.component_address, + value=update.value, + source_session_id=( + str(metadata_session_id) if metadata_session_id is not None else None + ), + source_process_id=source_process_id, + timestamp=update.timestamp, + ) + + return Command.COMPLETE.value + + async def _handle_process_route_response_request( + self, process_client_id: UUID, reader: asyncio.StreamReader + ) -> None: + response = await self._read_typed_payload( + reader, + ProcessControlResponse, + log_prefix=f"Process control {process_client_id} route response", + ) + if response is None: + return + + async with self._command_lock: + pending = self._pending_process_requests.pop(response.request_id, None) + + if pending is None: + logger.warning( + "Process control %s returned unknown request_id: %s", + process_client_id, + response.request_id, + ) + return + + owner_process_id, response_fut = pending + if owner_process_id != process_client_id: + if not response_fut.done(): + response_fut.set_result( + ProcessControlResponse( + request_id=response.request_id, + ok=False, + error=( + "Received response from unexpected process " + f"{process_client_id}; expected {owner_process_id}" + ), + error_code=ProcessControlErrorCode.INVALID_RESPONSE, + process_id=response.process_id, + ) + ) + return + + if not response_fut.done(): + response_fut.set_result(response) + + def _process_for_unit(self, unit_address: str) -> ProcessInfo | None: + for info in self.clients.values(): + if isinstance(info, ProcessInfo) and unit_address in info.units: + return info + return None + + def _process_owner_for_unit(self, unit_address: str) -> UUID | None: + for client_id, info in self.clients.items(): + if isinstance(info, ProcessInfo) and unit_address in info.units: + return client_id + return None + + def _metadata_collisions( + self, session_id: UUID, metadata: GraphMetadata + ) -> list[str]: + collisions: list[str] = [] + requested = set(metadata.components.keys()) + if not requested: + return collisions + for other_session_id, info in self.clients.items(): + if other_session_id == session_id or not isinstance(info, SessionInfo): + continue + if info.metadata is None: + continue + overlap = requested.intersection(info.metadata.components.keys()) + collisions.extend(overlap) + return sorted(set(collisions)) + + async def _route_process_request( + self, + unit_address: str, + operation: str, + payload: bytes | None, + timeout: float, + ) -> ProcessControlResponse: + request_id = str(uuid1()) + response_fut: asyncio.Future[ProcessControlResponse] = ( + asyncio.get_running_loop().create_future() + ) + request = ProcessControlRequest( + request_id=request_id, + unit_address=unit_address, + operation=operation, + payload=payload, + ) + + async with self._command_lock: + process_info = self._process_for_unit(unit_address) + if process_info is None: + return ProcessControlResponse( + request_id=request_id, + ok=False, + error=f"No process owns unit '{unit_address}'", + error_code=ProcessControlErrorCode.UNROUTABLE_UNIT, + ) + + self._pending_process_requests[request_id] = (process_info.id, response_fut) + + try: + async with process_info.write_lock: + process_writer = process_info.writer + request_bytes = pickle.dumps(request) + process_writer.write(Command.PROCESS_ROUTE_REQUEST.value) + process_writer.write(uint64_to_bytes(len(request_bytes))) + process_writer.write(request_bytes) + await process_writer.drain() + except Exception as exc: + self._pending_process_requests.pop(request_id, None) + return ProcessControlResponse( + request_id=request_id, + ok=False, + error=f"Failed to route request to owning process: {exc}", + error_code=ProcessControlErrorCode.ROUTE_WRITE_FAILED, + process_id=self._process_key(process_info.id), + ) + + try: + return await asyncio.wait_for(response_fut, timeout=timeout) + except asyncio.TimeoutError: + async with self._command_lock: + self._pending_process_requests.pop(request_id, None) + return ProcessControlResponse( + request_id=request_id, + ok=False, + error=( + f"Timed out waiting for process response " + f"(unit={unit_address}, operation={operation}, timeout={timeout}s)" + ), + error_code=ProcessControlErrorCode.TIMEOUT, + process_id=self._process_key(process_info.id), + ) + + async def _handle_session_process_request( + self, + writer: asyncio.StreamWriter, + reader: asyncio.StreamReader, + ) -> None: + unit_address = await read_str(reader) + operation = await read_str(reader) + timeout = float(await read_str(reader)) + payload_size = await read_int(reader) + payload: bytes | None = None + if payload_size > 0: + payload = await reader.readexactly(payload_size) + + response = await self._route_process_request( + unit_address=unit_address, + operation=operation, + payload=payload, + timeout=timeout, + ) + response_bytes = pickle.dumps(response) + writer.write(uint64_to_bytes(len(response_bytes))) + writer.write(response_bytes) + writer.write(Command.COMPLETE.value) + + def _connect_owner( + self, from_topic: str, to_topic: str, owner: UUID | str | None + ) -> bool: + topology_changed = self.graph.add_edge(from_topic, to_topic, owner=owner) + if isinstance(owner, UUID): + session = self._session_info(owner) + if session is not None: + session.edges.add((from_topic, to_topic)) + return topology_changed + + def _disconnect_owner( + self, from_topic: str, to_topic: str, owner: UUID | str | None + ) -> bool: + topology_changed = self.graph.remove_edge(from_topic, to_topic, owner=owner) + if isinstance(owner, UUID): + session = self._session_info(owner) + if session is not None: + session.edges.discard((from_topic, to_topic)) + return topology_changed + + def _session_info(self, session_id: UUID) -> SessionInfo | None: + info = self.clients.get(session_id) + if isinstance(info, SessionInfo): + return info + return None + + def _append_settings_event_locked( + self, + event_type: SettingsEventType, + component_address: str, + value: SettingsSnapshotValue, + source_session_id: str | None, + source_process_id: UUID | None, + timestamp: float | None = None, + ) -> None: + self._settings_event_seq += 1 + event = SettingsChangedEvent( + seq=self._settings_event_seq, + event_type=event_type, + component_address=component_address, + timestamp=timestamp if timestamp is not None else time.time(), + source_session_id=source_session_id, + source_process_id=source_process_id, + value=value, + ) + self._settings_events.append(event) + + for queue in self._settings_subscribers.values(): + self._queue_stream_event(queue, event) + + # Bound memory growth for long-lived servers. + max_events = self._retention_policy.settings_event_history_limit + if len(self._settings_events) > max_events: + del self._settings_events[0 : len(self._settings_events) - max_events] + + def _append_topology_event_locked( + self, + event_type: TopologyEventType, + changed_topics: list[str], + source_session_id: str | None, + source_process_id: UUID | None, + timestamp: float | None = None, + ) -> None: + self._topology_event_seq += 1 + event = TopologyChangedEvent( + seq=self._topology_event_seq, + event_type=event_type, + timestamp=timestamp if timestamp is not None else time.time(), + changed_topics=sorted(set(changed_topics)), + source_session_id=source_session_id, + source_process_id=source_process_id, + ) + self._topology_events.append(event) + + for queue in self._topology_subscribers.values(): + self._queue_stream_event(queue, event) + + max_events = self._retention_policy.topology_event_history_limit + if len(self._topology_events) > max_events: + del self._topology_events[0 : len(self._topology_events) - max_events] + + def _remove_settings_for_session_locked(self, session_id: UUID) -> None: + component_addresses = self._settings_owned_by_session.pop(session_id, set()) + for component_address in component_addresses: + state = self._settings_state.get(component_address) + if state is None or state.metadata_session_id != session_id: + continue + if state.source_process_id is None: + self._settings_state.pop(component_address, None) + else: + state.metadata_session_id = None + + def _session_owner_for_component_locked(self, component_address: str) -> UUID | None: + for client_id, info in self.clients.items(): + if not isinstance(info, SessionInfo): + continue + if info.metadata is None: + continue + if component_address in info.metadata.components: + return client_id + return None + + def _initial_settings_for_component_locked( + self, session_id: UUID, component_address: str + ) -> SettingsSnapshotValue | None: + session = self._session_info(session_id) + if session is None or session.metadata is None: + return None + component = session.metadata.components.get(component_address) + if component is None: + return None + initial_repr = component.initial_settings[1] + return SettingsSnapshotValue( + serialized=component.initial_settings[0], + repr_value=initial_repr, + structured_value=initial_repr if isinstance(initial_repr, dict) else None, + settings_schema=component.settings_schema, + ) + + def _remove_settings_for_process_locked(self, process_client_id: UUID) -> None: + source_process_id = self._process_key(process_client_id) + component_addresses = [ + component_address + for component_address, state in self._settings_state.items() + if state.source_process_id == source_process_id + ] + + for component_address in component_addresses: + state = self._settings_state.get(component_address) + if state is None: + continue + metadata_session_id = state.metadata_session_id + if metadata_session_id is None: + self._settings_state.pop(component_address, None) + continue + + restored = self._initial_settings_for_component_locked( + metadata_session_id, component_address + ) + if restored is None: + self._settings_state.pop(component_address, None) + continue + + state.value = restored + state.source_process_id = None + self._append_settings_event_locked( + event_type=SettingsEventType.SETTINGS_UPDATED, + component_address=component_address, + value=restored, + source_session_id=str(metadata_session_id), + source_process_id=None, + ) + + def _apply_session_metadata_settings_locked( + self, session_id: UUID, metadata: GraphMetadata + ) -> None: + session_components: set[str] = set() + for component in metadata.components.values(): + initial_repr = component.initial_settings[1] + value = SettingsSnapshotValue( + serialized=component.initial_settings[0], + repr_value=initial_repr, + structured_value=initial_repr if isinstance(initial_repr, dict) else None, + settings_schema=component.settings_schema, + ) + existing_state = self._settings_state.get(component.address) + if existing_state is not None and existing_state.source_process_id is not None: + existing_state.metadata_session_id = session_id + else: + self._settings_state[component.address] = _SettingsState( + value=value, + metadata_session_id=session_id, + source_process_id=None, + ) + session_components.add(component.address) + self._append_settings_event_locked( + event_type=SettingsEventType.INITIAL_SETTINGS, + component_address=component.address, + value=value, + source_session_id=str(session_id), + source_process_id=None, + ) + + self._settings_owned_by_session[session_id] = session_components + + async def _handle_session_edge_request( + self, + session_id: UUID, + req: bytes, + reader: asyncio.StreamReader, + ) -> bytes: + from_topic = await read_str(reader) + to_topic = await read_str(reader) + should_notify = False + + async with self._command_lock: + try: + if req == Command.SESSION_CONNECT.value: + should_notify = self._connect_owner( + from_topic, to_topic, session_id + ) + else: + should_notify = self._disconnect_owner( + from_topic, to_topic, session_id + ) + except CyclicException: + return Command.CYCLIC.value + + if should_notify: + self._append_topology_event_locked( + event_type=TopologyEventType.GRAPH_CHANGED, + changed_topics=[to_topic], + source_session_id=str(session_id), + source_process_id=None, + ) + + if should_notify: + await self._notify_downstream_for_topic(to_topic) + + return Command.COMPLETE.value + + async def _handle_session_clear_request(self, session_id: UUID) -> bytes: + async with self._command_lock: + notify_topics = self._clear_session_state(session_id) + + for topic in notify_topics: + await self._notify_downstream_for_topic(topic) + return Command.COMPLETE.value + + async def _handle_session_register_request( + self, session_id: UUID, reader: asyncio.StreamReader + ) -> bytes: + metadata = await self._read_typed_payload( + reader, + GraphMetadata, + log_prefix=f"Session {session_id} metadata", + ) + + async with self._command_lock: + session = self._session_info(session_id) + if session is not None and metadata is not None: + collisions = self._metadata_collisions(session_id, metadata) + if collisions: + logger.warning( + "Session %s metadata registration rejected due to component address collision(s): %s", + session_id, + ", ".join(collisions), + ) + return Command.ERROR.value + self._remove_settings_for_session_locked(session_id) + session.metadata = metadata + self._apply_session_metadata_settings_locked(session_id, metadata) + + return Command.COMPLETE.value + + async def _handle_session_snapshot_request( + self, writer: asyncio.StreamWriter + ) -> None: + async with self._command_lock: + snapshot = self._snapshot() + snapshot_bytes = pickle.dumps(snapshot) + writer.write(uint64_to_bytes(len(snapshot_bytes))) + writer.write(snapshot_bytes) + writer.write(Command.COMPLETE.value) + + async def _handle_session_settings_snapshot_request( + self, writer: asyncio.StreamWriter + ) -> None: + async with self._command_lock: + snapshot = { + component_address: self._settings_state[component_address].value + for component_address in sorted(self._settings_state) + } + snapshot_bytes = pickle.dumps(snapshot) + writer.write(uint64_to_bytes(len(snapshot_bytes))) + writer.write(snapshot_bytes) + writer.write(Command.COMPLETE.value) + + async def _handle_session_settings_events_request( + self, writer: asyncio.StreamWriter, after_seq: int + ) -> None: + async with self._command_lock: + events = [event for event in self._settings_events if event.seq > after_seq] + event_bytes = pickle.dumps(events) + writer.write(uint64_to_bytes(len(event_bytes))) + writer.write(event_bytes) + writer.write(Command.COMPLETE.value) + + def _clear_session_state(self, session_id: UUID) -> set[str]: + notify_topics: set[str] = set() + session = self._session_info(session_id) + if session is None: + return notify_topics + + for from_topic, to_topic in list(session.edges): + if self._disconnect_owner(from_topic, to_topic, session_id): + notify_topics.add(to_topic) + + self._remove_settings_for_session_locked(session_id) + session.metadata = None + if notify_topics: + self._append_topology_event_locked( + event_type=TopologyEventType.GRAPH_CHANGED, + changed_topics=list(notify_topics), + source_session_id=str(session_id), + source_process_id=None, + ) + return notify_topics + + def _drop_session(self, session_id: UUID) -> set[str]: + notify_topics: set[str] = set() + session = self._session_info(session_id) + if session is None: + return notify_topics + + for from_topic, to_topic in list(session.edges): + if self._disconnect_owner(from_topic, to_topic, session_id): + notify_topics.add(to_topic) + + self._remove_settings_for_session_locked(session_id) + session.metadata = None + self.clients.pop(session_id, None) + if notify_topics: + self._append_topology_event_locked( + event_type=TopologyEventType.GRAPH_CHANGED, + changed_topics=list(notify_topics), + source_session_id=str(session_id), + source_process_id=None, + ) + return notify_topics + + def _snapshot(self) -> GraphSnapshot: + graph = {node: sorted(conns) for node, conns in self.graph.graph.items()} + edge_owners = { + Edge(from_topic=from_topic, to_topic=to_topic): [ + "persistent" if owner is None else str(owner) + for owner in sorted( + owners, key=lambda owner: "" if owner is None else str(owner) + ) + ] + for (from_topic, to_topic), owners in sorted(self.graph.edge_owners.items()) + } + sessions = { + str(session_id): SnapshotSession( + edges=sorted( + [ + Edge(from_topic=from_topic, to_topic=to_topic) + for from_topic, to_topic in session.edges + ], + key=lambda edge: (edge.from_topic, edge.to_topic), + ), + metadata=session.metadata, + ) + for session_id, session in sorted( + [ + (client_id, info) + for client_id, info in self.clients.items() + if isinstance(info, SessionInfo) + ], + key=lambda item: str(item[0]), + ) + } + processes = { + client_id: SnapshotProcess( + process_id=self._process_key(client_id), + pid=process.pid, + host=process.host, + units=sorted(process.units), + ) + for client_id, process in sorted( + [ + (client_id, info) + for client_id, info in self.clients.items() + if isinstance(info, ProcessInfo) + ], + key=lambda item: str(item[0]), + ) + } + return GraphSnapshot( + graph=graph, + edge_owners=edge_owners, + sessions=sessions, + processes=processes, + ) + + async def _notify_downstream_for_topic(self, topic: str) -> None: + for sub in self._downstream_subs(topic): + await self._notify_subscriber(sub) + async def _notify_subscriber(self, sub: SubscriberInfo) -> None: try: pub_ids = [str(pub.id) for pub in self._upstream_pubs(sub.topic)] # Update requires us to read a 'COMPLETE' # This cannot be done from this context - async with sub.sync_writer() as writer: - notify_str = ",".join(pub_ids) - writer.write(Command.UPDATE.value) - writer.write(encode_str(notify_str)) + async def send_update() -> None: + async with sub.sync_writer() as writer: + notify_str = ",".join(pub_ids) + writer.write(Command.UPDATE.value) + writer.write(encode_str(notify_str)) + await asyncio.wait_for( + send_update(), timeout=SUBSCRIBER_UPDATE_TIMEOUT_SEC + ) except (ConnectionResetError, BrokenPipeError) as e: logger.debug(f"Failed to update Subscriber {sub.id}: {e}") + except asyncio.TimeoutError: + logger.warning( + "Timed out waiting for Subscriber %s to apply routing update for topic %s", + sub.id, + sub.topic, + ) def _publishers(self) -> list[PublisherInfo]: return [ diff --git a/src/ezmsg/core/messagechannel.py b/src/ezmsg/core/messagechannel.py index 8b50e298..130895f1 100644 --- a/src/ezmsg/core/messagechannel.py +++ b/src/ezmsg/core/messagechannel.py @@ -21,6 +21,7 @@ encode_str, close_stream_writer, ) +from .graphmeta import ProfileChannelType logger = logging.getLogger("ezmsg") @@ -99,6 +100,7 @@ class Channel: _pub_writer: asyncio.StreamWriter _graph_address: AddressType | None _local_backpressure: Backpressure | None + _channel_kind: ProfileChannelType def __init__( self, @@ -125,6 +127,7 @@ def __init__( self.clients = dict() self._graph_address = graph_address self._local_backpressure = None + self._channel_kind = ProfileChannelType.UNKNOWN @classmethod async def create( @@ -257,8 +260,10 @@ async def _publisher_connection(self, reader: asyncio.StreamReader) -> None: msg_id = await read_int(reader) buf_idx = msg_id % self.num_buffers + channel_kind = ProfileChannelType.UNKNOWN if msg == Command.TX_SHM.value: + channel_kind = ProfileChannelType.SHM shm_name = await read_str(reader) if self.shm is not None and self.shm.name != shm_name: @@ -285,6 +290,7 @@ async def _publisher_connection(self, reader: asyncio.StreamReader) -> None: self.cache.put_from_mem(self.shm[buf_idx]) elif msg == Command.TX_TCP.value: + channel_kind = ProfileChannelType.TCP buf_size = await read_int(reader) obj_bytes = await reader.readexactly(buf_size) assert MessageMarshal.msg_id(obj_bytes) == msg_id @@ -293,6 +299,8 @@ async def _publisher_connection(self, reader: asyncio.StreamReader) -> None: else: raise ValueError(f"unimplemented data telemetry: {msg}") + self._set_channel_kind(channel_kind) + if not self._notify_clients(msg_id): # Nobody is listening; need to ack! self.cache.release(msg_id) @@ -310,6 +318,22 @@ async def _publisher_connection(self, reader: asyncio.StreamReader) -> None: logger.debug(f"disconnected: channel:{self.id} -> pub:{self.pub_id}") + def _set_channel_kind(self, kind: ProfileChannelType) -> None: + if self._channel_kind == ProfileChannelType.UNKNOWN: + self._channel_kind = kind + elif self._channel_kind != kind: + logger.warning( + "Channel %s observed channel kind change: %s -> %s", + self.id, + self._channel_kind.value, + kind.value, + ) + self._channel_kind = kind + + @property + def channel_kind(self) -> ProfileChannelType: + return self._channel_kind + def _notify_clients(self, msg_id: int) -> bool: """notify interested clients and return true if any were notified""" buf_idx = msg_id % self.num_buffers @@ -331,6 +355,7 @@ def put_local(self, msg_id: int, msg: typing.Any) -> None: ) buf_idx = msg_id % self.num_buffers + self._set_channel_kind(ProfileChannelType.LOCAL) if self._notify_clients(msg_id): self.cache.put_local(msg, msg_id) self._local_backpressure.lease(self.id, buf_idx) diff --git a/src/ezmsg/core/netprotocol.py b/src/ezmsg/core/netprotocol.py index 04ab0839..ff6bf242 100644 --- a/src/ezmsg/core/netprotocol.py +++ b/src/ezmsg/core/netprotocol.py @@ -9,6 +9,7 @@ from dataclasses import field, dataclass from contextlib import asynccontextmanager from asyncio.base_events import Server +from .graphmeta import GraphMetadata VERSION = b"1" UINT64_SIZE = 8 @@ -165,6 +166,28 @@ class ChannelInfo(ClientInfo): pub_id: UUID +@dataclass +class SessionInfo(ClientInfo): + """ + Session-scoped control-plane client information. + """ + + edges: set[tuple[str, str]] = field(default_factory=set) + metadata: GraphMetadata | None = None + + +@dataclass +class ProcessInfo(ClientInfo): + """ + Process-scoped control-plane client information. + """ + + pid: int | None = None + host: str | None = None + units: set[str] = field(default_factory=set) + write_lock: asyncio.Lock = field(default_factory=asyncio.Lock, init=False) + + def uint64_to_bytes(i: int) -> bytes: """ Convert a 64-bit unsigned integer to bytes. @@ -300,6 +323,30 @@ def _generate_next_value_(name, start, count, last_values) -> bytes: SHM_OK = enum.auto() SHM_ATTACH_FAILED = enum.auto() + # GraphContext Session Commands (control plane) + SESSION = enum.auto() + SESSION_CONNECT = enum.auto() + SESSION_DISCONNECT = enum.auto() + SESSION_CLEAR = enum.auto() + SESSION_REGISTER = enum.auto() + SESSION_SNAPSHOT = enum.auto() + SESSION_SETTINGS_SNAPSHOT = enum.auto() + SESSION_SETTINGS_EVENTS = enum.auto() + SESSION_SETTINGS_SUBSCRIBE = enum.auto() + SESSION_TOPOLOGY_SUBSCRIBE = enum.auto() + SESSION_PROFILING_SUBSCRIBE = enum.auto() + SESSION_PROCESS_REQUEST = enum.auto() + + # Backend Process Control Commands + PROCESS = enum.auto() + PROCESS_REGISTER = enum.auto() + PROCESS_UPDATE_OWNERSHIP = enum.auto() + PROCESS_SETTINGS_UPDATE = enum.auto() + PROCESS_PROFILING_TRACE_UPDATE = enum.auto() + PROCESS_ROUTE_REQUEST = enum.auto() + PROCESS_ROUTE_RESPONSE = enum.auto() + ERROR = enum.auto() + def create_socket( host: str | None = None, diff --git a/src/ezmsg/core/processclient.py b/src/ezmsg/core/processclient.py new file mode 100644 index 00000000..97c12cec --- /dev/null +++ b/src/ezmsg/core/processclient.py @@ -0,0 +1,483 @@ +import asyncio +import logging +import os +import pickle +import socket +import time + +from uuid import UUID +from contextlib import suppress +from collections.abc import Awaitable, Callable + +from .graphmeta import ( + ProcessProfilingSnapshot, + ProcessProfilingTraceBatch, + ProcessControlErrorCode, + ProcessControlOperation, + ProcessControlRequest, + ProcessControlResponse, + ProfilingTraceControl, + ProcessPing, + ProcessRegistration, + ProcessStats, + ProcessOwnershipUpdate, + ProcessSettingsUpdate, + SettingsSnapshotValue, +) +from .profiling import PROFILES +from .graphserver import GraphService +from .netprotocol import ( + AddressType, + Command, + close_stream_writer, + read_int, + read_str, + uint64_to_bytes, +) + +logger = logging.getLogger("ezmsg") + + +class ProcessControlClient: + _graph_address: AddressType | None + _client_id: UUID | None + _reader: asyncio.StreamReader | None + _writer: asyncio.StreamWriter | None + _write_lock: asyncio.Lock + _ack_queue: asyncio.Queue[bytes] + _io_task: asyncio.Task[None] | None + _request_handler: Callable[ + [ProcessControlRequest], ProcessControlResponse | Awaitable[ProcessControlResponse] + ] | None + _owned_units: set[str] + _trace_push_task: asyncio.Task[None] | None + _trace_push_interval_s: float + _trace_push_max_samples: int + + def __init__(self, graph_address: AddressType | None = None) -> None: + self._graph_address = graph_address + self._client_id = None + self._reader = None + self._writer = None + self._write_lock = asyncio.Lock() + self._ack_queue = asyncio.Queue() + self._io_task = None + self._request_handler = None + self._owned_units = set() + self._trace_push_task = None + self._trace_push_interval_s = float( + os.environ.get("EZMSG_PROFILE_TRACE_PUSH_INTERVAL_S", "0.02") + ) + self._trace_push_max_samples = int( + os.environ.get("EZMSG_PROFILE_TRACE_PUSH_MAX_SAMPLES", "5000") + ) + + def _require_client_id(self) -> UUID: + if self._client_id is None: + raise RuntimeError("Process control connection is not active") + return self._client_id + + @property + def process_id(self) -> UUID: + return self._require_client_id() + + @property + def client_id(self) -> UUID | None: + return self._client_id + + async def connect(self) -> None: + if self._writer is not None: + return + + reader, writer = await GraphService(self._graph_address).open_connection() + writer.write(Command.PROCESS.value) + await writer.drain() + + client_id = UUID(await read_str(reader)) + response = await reader.read(1) + if response != Command.COMPLETE.value: + await close_stream_writer(writer) + raise RuntimeError("Failed to create process control connection") + + self._client_id = client_id + PROFILES.set_process_id(client_id) + self._reader = reader + self._writer = writer + self._io_task = asyncio.create_task( + self._io_loop(), + name=f"process-control-{client_id}", + ) + + def set_request_handler( + self, + handler: Callable[ + [ProcessControlRequest], ProcessControlResponse | Awaitable[ProcessControlResponse] + ] + | None, + ) -> None: + self._request_handler = handler + + async def register(self, units: list[str]) -> None: + await self.connect() + normalized_units = sorted(set(units)) + payload = ProcessRegistration( + pid=os.getpid(), + host=socket.gethostname(), + units=normalized_units, + ) + await self._payload_command(Command.PROCESS_REGISTER, payload) + self._owned_units = set(normalized_units) + + async def update_ownership( + self, + added_units: list[str] | None = None, + removed_units: list[str] | None = None, + ) -> None: + await self.connect() + added = sorted(set(added_units or [])) + removed = sorted(set(removed_units or [])) + payload = ProcessOwnershipUpdate( + added_units=added, + removed_units=removed, + ) + await self._payload_command(Command.PROCESS_UPDATE_OWNERSHIP, payload) + self._owned_units.update(added) + self._owned_units.difference_update(removed) + + async def report_settings_update( + self, + component_address: str, + value: SettingsSnapshotValue, + timestamp: float | None = None, + ) -> None: + await self.connect() + payload = ProcessSettingsUpdate( + component_address=component_address, + value=value, + timestamp=timestamp if timestamp is not None else time.time(), + ) + await self._payload_command(Command.PROCESS_SETTINGS_UPDATE, payload) + + async def close(self) -> None: + writer = self._writer + if writer is None: + return + + trace_task = self._trace_push_task + self._trace_push_task = None + if trace_task is not None: + trace_task.cancel() + with suppress(asyncio.CancelledError): + await trace_task + + io_task = self._io_task + self._io_task = None + if io_task is not None: + io_task.cancel() + with suppress(asyncio.CancelledError): + await io_task + + self._reader = None + self._writer = None + self._client_id = None + await close_stream_writer(writer) + + async def _payload_command(self, command: Command, payload_obj: object) -> None: + await self._write_payload(command, payload_obj, expect_complete=True) + + async def _write_payload( + self, + command: Command, + payload_obj: object, + *, + expect_complete: bool, + ) -> None: + reader = self._reader + writer = self._writer + if reader is None or writer is None: + raise RuntimeError("Process control connection is not active") + + payload = pickle.dumps(payload_obj) + async with self._write_lock: + writer.write(command.value) + writer.write(uint64_to_bytes(len(payload))) + writer.write(payload) + await writer.drain() + + if not expect_complete: + return + + try: + response = await asyncio.wait_for(self._ack_queue.get(), timeout=5.0) + except asyncio.TimeoutError as exc: + raise RuntimeError( + f"Timed out waiting for response to process control command: {command.name}" + ) from exc + + if response != Command.COMPLETE.value: + if response == Command.ERROR.value: + raise RuntimeError( + f"Process control command failed: {command.name}" + ) + raise RuntimeError( + f"Unexpected response to process control command: {command.name}" + ) + + async def _io_loop(self) -> None: + reader = self._reader + writer = self._writer + if reader is None or writer is None: + return + + try: + while True: + req = await reader.read(1) + if not req: + break + + if req in (Command.COMPLETE.value, Command.ERROR.value): + self._ack_queue.put_nowait(req) + continue + + if req != Command.PROCESS_ROUTE_REQUEST.value: + logger.warning( + "Process control %s received unknown command: %s", + self._client_id, + req, + ) + continue + + payload_size = await read_int(reader) + payload = await reader.readexactly(payload_size) + request: ProcessControlRequest | None = None + try: + request_obj = pickle.loads(payload) + if isinstance(request_obj, ProcessControlRequest): + request = request_obj + else: + raise RuntimeError( + "process route request payload was not ProcessControlRequest" + ) + except Exception as exc: + logger.warning( + "Process control %s failed to parse route request: %s", + self._client_id, + exc, + ) + + if request is None: + continue + + response = await self._handle_route_request(request) + await self._write_payload( + Command.PROCESS_ROUTE_RESPONSE, + response, + expect_complete=False, + ) + + except asyncio.CancelledError: + raise + except (ConnectionResetError, BrokenPipeError) as exc: + logger.debug(f"Process control {self._client_id} disconnected: {exc}") + + async def _handle_route_request( + self, request: ProcessControlRequest + ) -> ProcessControlResponse: + operation: ProcessControlOperation | None = None + if isinstance(request.operation, ProcessControlOperation): + operation = request.operation + elif isinstance(request.operation, str): + with suppress(ValueError): + operation = ProcessControlOperation(request.operation) + + if operation == ProcessControlOperation.PING: + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + payload=pickle.dumps( + ProcessPing( + process_id=self.process_id, + pid=os.getpid(), + host=socket.gethostname(), + timestamp=time.time(), + ) + ), + process_id=self.process_id, + ) + + if operation == ProcessControlOperation.GET_PROCESS_STATS: + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + payload=pickle.dumps( + ProcessStats( + process_id=self.process_id, + pid=os.getpid(), + host=socket.gethostname(), + owned_units=sorted(self._owned_units), + timestamp=time.time(), + ) + ), + process_id=self.process_id, + ) + + if operation == ProcessControlOperation.GET_PROFILING_SNAPSHOT: + snapshot: ProcessProfilingSnapshot = PROFILES.snapshot() + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + payload=pickle.dumps(snapshot), + process_id=self.process_id, + ) + + if operation == ProcessControlOperation.SET_PROFILING_TRACE: + control: ProfilingTraceControl | None = None + try: + if request.payload is not None: + control_obj = pickle.loads(request.payload) + if isinstance(control_obj, ProfilingTraceControl): + control = control_obj + except Exception as exc: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=f"Invalid profiling trace control payload: {exc}", + error_code=ProcessControlErrorCode.INVALID_RESPONSE, + process_id=self.process_id, + ) + + if control is None: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error="Missing profiling trace control payload", + error_code=ProcessControlErrorCode.INVALID_RESPONSE, + process_id=self.process_id, + ) + + PROFILES.set_trace_control(control) + if control.enabled: + await self._ensure_trace_push_task() + else: + await self._cancel_trace_push_task() + + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + process_id=self.process_id, + ) + + if operation == ProcessControlOperation.GET_PROFILING_TRACE_BATCH: + max_samples = 1000 + if request.payload is not None: + try: + max_samples_obj = pickle.loads(request.payload) + if isinstance(max_samples_obj, int): + max_samples = max(1, max_samples_obj) + except Exception: + pass + + batch: ProcessProfilingTraceBatch = PROFILES.trace_batch( + max_samples=max_samples + ) + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + payload=pickle.dumps(batch), + process_id=self.process_id, + ) + + if self._request_handler is None: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=f"Unsupported process control operation: {request.operation}", + error_code=ProcessControlErrorCode.HANDLER_NOT_CONFIGURED, + process_id=self.process_id, + ) + + try: + result = self._request_handler(request) + if asyncio.iscoroutine(result): + result = await result + except Exception as exc: + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=f"process request handler failed: {exc}", + error_code=ProcessControlErrorCode.HANDLER_ERROR, + process_id=self.process_id, + ) + + if not isinstance(result, ProcessControlResponse): + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=( + "process request handler returned invalid response type: " + f"{type(result).__name__}" + ), + error_code=ProcessControlErrorCode.INVALID_RESPONSE, + process_id=self.process_id, + ) + + if result.request_id != request.request_id: + result = ProcessControlResponse( + request_id=request.request_id, + ok=False, + error=( + "process request handler returned mismatched request_id: " + f"{result.request_id}" + ), + error_code=ProcessControlErrorCode.INVALID_RESPONSE, + process_id=self.process_id, + ) + + if result.process_id is None: + result.process_id = self.process_id + + return result + + async def _ensure_trace_push_task(self) -> None: + task = self._trace_push_task + if task is not None and not task.done(): + return + self._trace_push_task = asyncio.create_task( + self._trace_push_loop(), + name=f"proc-trace-push-{self.process_id}", + ) + + async def _cancel_trace_push_task(self) -> None: + task = self._trace_push_task + self._trace_push_task = None + if task is None: + return + task.cancel() + with suppress(asyncio.CancelledError): + await task + + async def _trace_push_loop(self) -> None: + try: + while True: + await asyncio.sleep(max(0.01, self._trace_push_interval_s)) + batch: ProcessProfilingTraceBatch = PROFILES.trace_batch( + max_samples=max(1, self._trace_push_max_samples) + ) + if len(batch.samples) > 0: + await self._write_payload( + Command.PROCESS_PROFILING_TRACE_UPDATE, + batch, + expect_complete=False, + ) + + if not PROFILES.trace_enabled(): + break + except asyncio.CancelledError: + raise + except (ConnectionResetError, BrokenPipeError): + logger.debug("Process trace push loop disconnected") + except Exception as exc: + logger.warning(f"Process trace push loop failed: {exc}") + finally: + if asyncio.current_task() is self._trace_push_task: + self._trace_push_task = None diff --git a/src/ezmsg/core/profiling.py b/src/ezmsg/core/profiling.py new file mode 100644 index 00000000..27b9cdf8 --- /dev/null +++ b/src/ezmsg/core/profiling.py @@ -0,0 +1,445 @@ +import os +import socket +import time +import heapq + +from collections import deque +from dataclasses import dataclass, field +from typing import Callable, TypeAlias +from uuid import UUID + +from .graphmeta import ( + ProcessProfilingSnapshot, + ProcessProfilingTraceBatch, + ProfileChannelType, + ProfilingTraceControl, + ProfilingTraceSample, + PublisherProfileSnapshot, + SubscriberProfileSnapshot, +) + + +TRACE_MAX_SAMPLES = int(os.environ.get("EZMSG_PROFILE_TRACE_MAX_SAMPLES", "10000")) +# Must return monotonic nanoseconds so *_ns metrics remain unit-consistent. +PROFILE_TIME_TYPE: TypeAlias = Callable[[], int] +PROFILE_TIME: PROFILE_TIME_TYPE = time.perf_counter_ns + + +def _endpoint_id(topic: str, id: UUID) -> str: + return f"{topic}:{id}" + + +@dataclass +class _PublisherMetrics: + topic: str + endpoint_id: str + num_buffers: int + messages_published_total: int = 0 + inflight_messages_current: int = 0 + _last_snapshot_total: int = 0 + _last_publish_ts_ns: int | None = None + trace_enabled: bool = False + trace_sample_mod: int = 1 + trace_metrics: set[str] | None = None + _trace_counter: int = 0 + _trace_publish_delta_enabled: bool = False + _trace_backpressure_wait_enabled: bool = False + trace_samples: deque[ProfilingTraceSample] = field( + default_factory=lambda: deque(maxlen=TRACE_MAX_SAMPLES) + ) + + def record_publish(self, inflight: int, msg_seq: int | None = None) -> None: + self.messages_published_total += 1 + self.inflight_messages_current = inflight + + if not self._trace_publish_delta_enabled: + return + self._trace_counter += 1 + if self._trace_counter % max(1, self.trace_sample_mod) != 0: + return + + now_ns = PROFILE_TIME() + publish_delta_ns = ( + 0 if self._last_publish_ts_ns is None else now_ns - self._last_publish_ts_ns + ) + self._last_publish_ts_ns = now_ns + self.trace_samples.append( + ProfilingTraceSample( + timestamp=float(now_ns), + endpoint_id=self.endpoint_id, + topic=self.topic, + metric="publish_delta_ns", + value=float(publish_delta_ns), + sample_seq=msg_seq, + ) + ) + + def record_backpressure_wait(self, wait_ns: int, msg_seq: int | None = None) -> None: + if not self._trace_backpressure_wait_enabled: + return + + now_ns = PROFILE_TIME() + self.trace_samples.append( + ProfilingTraceSample( + timestamp=float(now_ns), + endpoint_id=self.endpoint_id, + topic=self.topic, + metric="backpressure_wait_ns", + value=float(wait_ns), + sample_seq=msg_seq, + ) + ) + + def sample_inflight(self, inflight: int) -> None: + self.inflight_messages_current = inflight + + def snapshot( + self, + now_ns: int, + window_seconds: float, + *, + has_previous_snapshot: bool, + ) -> PublisherProfileSnapshot: + window_count = ( + self.messages_published_total - self._last_snapshot_total + if has_previous_snapshot + else 0 + ) + self._last_snapshot_total = self.messages_published_total + return PublisherProfileSnapshot( + endpoint_id=self.endpoint_id, + topic=self.topic, + messages_published_total=self.messages_published_total, + messages_published_window=window_count, + publish_rate_hz_window=( + float(window_count) / max(window_seconds, 1e-9) + if has_previous_snapshot and window_seconds > 0.0 + else 0.0 + ), + inflight_messages_current=self.inflight_messages_current, + num_buffers=self.num_buffers, + timestamp=float(now_ns), + ) + + +@dataclass +class _SubscriberMetrics: + topic: str + endpoint_id: str + messages_received_total: int = 0 + channel_kind_last: ProfileChannelType = ProfileChannelType.UNKNOWN + _last_snapshot_total: int = 0 + trace_enabled: bool = False + trace_sample_mod: int = 1 + trace_metrics: set[str] | None = None + _trace_counter: int = 0 + _trace_lease_time_enabled: bool = False + _trace_user_span_enabled: bool = False + trace_samples: deque[ProfilingTraceSample] = field( + default_factory=lambda: deque(maxlen=TRACE_MAX_SAMPLES) + ) + + def begin_message(self, channel_kind: ProfileChannelType) -> bool: + self.messages_received_total += 1 + self.channel_kind_last = channel_kind + + if not (self._trace_lease_time_enabled or self._trace_user_span_enabled): + return False + + self._trace_counter += 1 + return self._trace_counter % max(1, self.trace_sample_mod) == 0 + + def record_receive( + self, + channel_kind: ProfileChannelType, + lease_ns: int | None = None, + msg_seq: int | None = None, + ) -> None: + sampled = self.begin_message(channel_kind) + self.record_lease_time( + channel_kind, + lease_ns, + msg_seq=msg_seq, + sampled=sampled, + ) + + def record_lease_time( + self, + channel_kind: ProfileChannelType, + lease_ns: int | None, + msg_seq: int | None = None, + *, + sampled: bool, + ) -> None: + if lease_ns is None or not self._trace_lease_time_enabled or not sampled: + return + + now_ns = PROFILE_TIME() + self.trace_samples.append( + ProfilingTraceSample( + timestamp=float(now_ns), + endpoint_id=self.endpoint_id, + topic=self.topic, + metric="lease_time_ns", + value=float(lease_ns), + channel_kind=channel_kind, + sample_seq=msg_seq, + ) + ) + + def record_user_span( + self, + span_ns: int, + label: str | None, + msg_seq: int | None = None, + *, + sampled: bool, + ) -> None: + if not self._trace_user_span_enabled or not sampled: + return + + now_ns = PROFILE_TIME() + self.trace_samples.append( + ProfilingTraceSample( + timestamp=float(now_ns), + endpoint_id=self.endpoint_id, + topic=self.topic if label is None else f"{self.topic}:{label}", + metric="user_span_ns", + value=float(span_ns), + channel_kind=self.channel_kind_last, + sample_seq=msg_seq, + ) + ) + + def snapshot( + self, + now_ns: int, + *, + has_previous_snapshot: bool, + ) -> SubscriberProfileSnapshot: + window_count = ( + self.messages_received_total - self._last_snapshot_total + if has_previous_snapshot + else 0 + ) + self._last_snapshot_total = self.messages_received_total + return SubscriberProfileSnapshot( + endpoint_id=self.endpoint_id, + topic=self.topic, + messages_received_total=self.messages_received_total, + messages_received_window=window_count, + channel_kind_last=self.channel_kind_last, + timestamp=float(now_ns), + ) + + +class ProfileRegistry: + def __init__(self) -> None: + self._process_id = UUID(int=0) + self._pid = os.getpid() + self._host = socket.gethostname() + self._publishers: dict[UUID, _PublisherMetrics] = {} + self._subscribers: dict[UUID, _SubscriberMetrics] = {} + self._default_trace_control = ProfilingTraceControl(enabled=False) + self._trace_control_expires_ns: int | None = None + self._last_snapshot_ts_ns: int | None = None + + def set_process_id(self, process_id: UUID, *, reset: bool = False) -> None: + if reset: + self._publishers.clear() + self._subscribers.clear() + self._default_trace_control = ProfilingTraceControl(enabled=False) + self._trace_control_expires_ns = None + self._process_id = process_id + self._last_snapshot_ts_ns = None + + def register_publisher(self, pub_id: UUID, topic: str, num_buffers: int) -> _PublisherMetrics: + metric = _PublisherMetrics( + topic=topic, + endpoint_id=_endpoint_id(topic, pub_id), + num_buffers=max(1, int(num_buffers)), + ) + self._publishers[pub_id] = metric + self._apply_trace_control_to_publisher(metric) + return metric + + def unregister_publisher(self, pub_id: UUID) -> None: + self._publishers.pop(pub_id, None) + + def register_subscriber(self, sub_id: UUID, topic: str) -> _SubscriberMetrics: + metric = _SubscriberMetrics( + topic=topic, + endpoint_id=_endpoint_id(topic, sub_id), + ) + self._subscribers[sub_id] = metric + self._apply_trace_control_to_subscriber(metric) + return metric + + def unregister_subscriber(self, sub_id: UUID) -> None: + self._subscribers.pop(sub_id, None) + + def snapshot(self) -> ProcessProfilingSnapshot: + now_ns = PROFILE_TIME() + last_snapshot_ts_ns = self._last_snapshot_ts_ns + has_previous_snapshot = last_snapshot_ts_ns is not None + window_seconds = ( + float(now_ns - last_snapshot_ts_ns) / 1e9 + if has_previous_snapshot + else 0.0 + ) + self._last_snapshot_ts_ns = now_ns + return ProcessProfilingSnapshot( + process_id=self._process_id, + pid=self._pid, + host=self._host, + window_seconds=window_seconds, + timestamp=float(now_ns), + publishers={ + metric.endpoint_id: metric.snapshot( + now_ns, + window_seconds, + has_previous_snapshot=has_previous_snapshot, + ) + for metric in self._publishers.values() + }, + subscribers={ + metric.endpoint_id: metric.snapshot( + now_ns, + has_previous_snapshot=has_previous_snapshot, + ) + for metric in self._subscribers.values() + }, + ) + + def set_trace_control(self, control: ProfilingTraceControl) -> None: + # Changing filters/mode should start from a clean trace buffer so new + # consumers do not receive stale samples from an old control scope. + self._clear_trace_samples() + self._default_trace_control = control + if control.enabled and control.ttl_seconds is not None: + self._trace_control_expires_ns = PROFILE_TIME() + max( + 0, int(control.ttl_seconds * 1e9) + ) + else: + self._trace_control_expires_ns = None + + for metric in self._publishers.values(): + self._apply_trace_control_to_publisher(metric) + + for metric in self._subscribers.values(): + self._apply_trace_control_to_subscriber(metric) + + def trace_batch(self, max_samples: int = 1000) -> ProcessProfilingTraceBatch: + self._expire_trace_control_if_needed() + samples: list[ProfilingTraceSample] = [] + limit = max(1, int(max_samples)) + + queues: list[deque[ProfilingTraceSample]] = [] + for metric in self._publishers.values(): + if metric.trace_samples: + queues.append(metric.trace_samples) + for metric in self._subscribers.values(): + if metric.trace_samples: + queues.append(metric.trace_samples) + + if len(queues) == 1: + queue = queues[0] + while queue and len(samples) < limit: + samples.append(queue.popleft()) + elif len(queues) > 1: + heap: list[tuple[float, int, int]] = [] + for idx, queue in enumerate(queues): + sample = queue[0] + seq = sample.sample_seq if sample.sample_seq is not None else -1 + heapq.heappush(heap, (sample.timestamp, seq, idx)) + + while heap and len(samples) < limit: + _timestamp, _seq, queue_idx = heapq.heappop(heap) + queue = queues[queue_idx] + if not queue: + continue + sample = queue.popleft() + samples.append(sample) + if queue: + nxt = queue[0] + nxt_seq = nxt.sample_seq if nxt.sample_seq is not None else -1 + heapq.heappush(heap, (nxt.timestamp, nxt_seq, queue_idx)) + + return ProcessProfilingTraceBatch( + process_id=self._process_id, + pid=self._pid, + host=self._host, + timestamp=float(PROFILE_TIME()), + samples=samples, + ) + + def trace_enabled(self) -> bool: + self._expire_trace_control_if_needed() + return self._default_trace_control.enabled + + def _expire_trace_control_if_needed(self, now_ns: int | None = None) -> None: + expires_ns = self._trace_control_expires_ns + if expires_ns is None: + return + ts_ns = now_ns if now_ns is not None else PROFILE_TIME() + if ts_ns < expires_ns: + return + self.set_trace_control(ProfilingTraceControl(enabled=False)) + + def _apply_trace_control_to_publisher(self, metric: _PublisherMetrics) -> None: + control = self._default_trace_control + sample_mod = max(1, control.sample_mod) + pub_topics = set(control.publisher_topics or []) + pub_endpoint_ids = set(control.publisher_endpoint_ids or []) + trace_metrics = ( + set(control.metrics) if control.metrics is not None else None + ) + enabled = control.enabled + if enabled and pub_topics and metric.topic not in pub_topics: + enabled = False + if enabled and pub_endpoint_ids and metric.endpoint_id not in pub_endpoint_ids: + enabled = False + metric.trace_enabled = enabled + metric.trace_sample_mod = sample_mod + metric.trace_metrics = trace_metrics + metric._trace_counter = 0 + metric._last_publish_ts_ns = None + metric._trace_publish_delta_enabled = enabled and ( + trace_metrics is None or "publish_delta_ns" in trace_metrics + ) + metric._trace_backpressure_wait_enabled = enabled and ( + trace_metrics is None or "backpressure_wait_ns" in trace_metrics + ) + + def _apply_trace_control_to_subscriber(self, metric: _SubscriberMetrics) -> None: + control = self._default_trace_control + sample_mod = max(1, control.sample_mod) + sub_topics = set(control.subscriber_topics or []) + sub_endpoint_ids = set(control.subscriber_endpoint_ids or []) + trace_metrics = ( + set(control.metrics) if control.metrics is not None else None + ) + enabled = control.enabled + if enabled and sub_topics and metric.topic not in sub_topics: + enabled = False + if enabled and sub_endpoint_ids and metric.endpoint_id not in sub_endpoint_ids: + enabled = False + metric.trace_enabled = enabled + metric.trace_sample_mod = sample_mod + metric.trace_metrics = trace_metrics + metric._trace_counter = 0 + metric._trace_lease_time_enabled = enabled and ( + trace_metrics is None or "lease_time_ns" in trace_metrics + ) + metric._trace_user_span_enabled = enabled and ( + trace_metrics is None or "user_span_ns" in trace_metrics + ) + + def _clear_trace_samples(self) -> None: + for metric in self._publishers.values(): + metric.trace_samples.clear() + for metric in self._subscribers.values(): + metric.trace_samples.clear() + + +PROFILES = ProfileRegistry() diff --git a/src/ezmsg/core/pubclient.py b/src/ezmsg/core/pubclient.py index f9c42952..706c6c7a 100644 --- a/src/ezmsg/core/pubclient.py +++ b/src/ezmsg/core/pubclient.py @@ -13,6 +13,7 @@ from .channelmanager import CHANNELS from .messagechannel import Channel from .messagemarshal import MessageMarshal, UninitializedMemory +from .profiling import PROFILES, PROFILE_TIME from .netprotocol import ( Address, @@ -37,6 +38,36 @@ BACKPRESSURE_WARNING = "EZMSG_DISABLE_BACKPRESSURE_WARNING" not in os.environ BACKPRESSURE_REFRACTORY = 5.0 # sec +ALLOW_LOCAL_ENV = "EZMSG_ALLOW_LOCAL" +FORCE_TCP_ENV = "EZMSG_FORCE_TCP" + + +def _process_allow_local_default() -> bool: + value = os.environ.get(ALLOW_LOCAL_ENV, "") + if value == "": + return True + return value.lower() in ("1", "true", "yes", "on") + + +def _process_force_tcp_default() -> bool: + value = os.environ.get(FORCE_TCP_ENV, "") + if value == "": + return False + return value.lower() in ("1", "true", "yes", "on") + + +def _resolve_force_tcp(force_tcp: bool | None) -> bool: + if force_tcp is None: + return _process_force_tcp_default() + return force_tcp + + +def _resolve_allow_local(force_tcp: bool, allow_local: bool | None) -> bool: + resolved = _process_allow_local_default() if allow_local is None else allow_local + if force_tcp and resolved: + logger.info("force_tcp=True disables local delivery for this publisher") + return False + return resolved # Publisher needs a bit more information about connected channels @@ -75,7 +106,9 @@ class Publisher: _msg_id: int _shm: SHMContext _force_tcp: bool + _allow_local: bool _last_backpressure_event: float + _profile: object _graph_address: AddressType | None @@ -99,7 +132,8 @@ async def create( buf_size: int = DEFAULT_SHM_SIZE, num_buffers: int = 32, start_paused: bool = False, - force_tcp: bool = False, + force_tcp: bool | None = None, + allow_local: bool | None = None, ) -> "Publisher": """ Create a new Publisher instance and register it with the graph server. @@ -116,6 +150,16 @@ async def create( :type port: int | None :param buf_size: Size of shared memory buffers. :type buf_size: int + :param force_tcp: Whether to force TCP transport instead of shared memory. + If None, inherit the process default from ``EZMSG_FORCE_TCP`` which + defaults to disabled. + :type force_tcp: bool | None + :param allow_local: Whether to allow the in-process fast path when available. + If None, inherit the process default from ``EZMSG_ALLOW_LOCAL`` which + defaults to enabled. Set to False to bypass local delivery and + characterize same-process SHM or TCP. When ``force_tcp=True``, local + delivery is disabled regardless of this value. + :type allow_local: bool | None :param kwargs: Additional keyword arguments for Publisher constructor. :return: Initialized and registered Publisher instance. :rtype: Publisher @@ -127,6 +171,8 @@ async def create( writer.write(Command.PUBLISH.value) writer.write(encode_str(topic)) + resolved_force_tcp = _resolve_force_tcp(force_tcp) + pub_id = UUID(await read_str(reader)) pub = cls( id=pub_id, @@ -135,7 +181,8 @@ async def create( graph_address=graph_address, num_buffers=num_buffers, start_paused=start_paused, - force_tcp=force_tcp, + force_tcp=resolved_force_tcp, + allow_local=allow_local, _guard=cls._SENTINEL, ) @@ -189,7 +236,8 @@ def __init__( graph_address: AddressType | None = None, num_buffers: int = 32, start_paused: bool = False, - force_tcp: bool = False, + force_tcp: bool | None = None, + allow_local: bool | None = None, _guard = None ) -> None: """ @@ -207,7 +255,12 @@ def __init__( :param start_paused: Whether to start in paused state. :type start_paused: bool :param force_tcp: Whether to force TCP transport instead of shared memory. - :type force_tcp: bool + If None, inherit the process default from ``EZMSG_FORCE_TCP``. + :type force_tcp: bool | None + :param allow_local: Whether to allow the direct in-process fast path when available. + If None, inherit the process default from ``EZMSG_ALLOW_LOCAL``. + When ``force_tcp=True``, local delivery is disabled regardless of this value. + :type allow_local: bool | None """ if _guard is not self._SENTINEL: raise TypeError( @@ -227,9 +280,11 @@ def __init__( self._running.set() self._num_buffers = num_buffers self._backpressure = Backpressure(num_buffers) - self._force_tcp = force_tcp + self._force_tcp = _resolve_force_tcp(force_tcp) + self._allow_local = _resolve_allow_local(self._force_tcp, allow_local) self._last_backpressure_event = -1 self._graph_address = graph_address + self._profile = PROFILES.register_publisher(self.id, self.topic, self._num_buffers) @property def log_name(self) -> str: @@ -243,6 +298,7 @@ def close(self) -> None: and all subscriber handling tasks. """ self._graph_task.cancel() + PROFILES.unregister_publisher(self.id) self._shm.close() self._connection_task.cancel() for task in self._channel_tasks.values(): @@ -369,12 +425,14 @@ async def _handle_channel( elif msg == Command.RX_ACK.value: msg_id = await read_int(reader) self._backpressure.free(info.id, msg_id % self._num_buffers) + self._profile.sample_inflight(self._backpressure.pressure) except (ConnectionResetError, BrokenPipeError): logger.debug(f"Publisher {self.id}: Channel {info.id} connection fail") finally: self._backpressure.free(info.id) + self._profile.sample_inflight(self._backpressure.pressure) await close_stream_writer(self._channels[info.id].writer) del self._channels[info.id] @@ -434,14 +492,18 @@ async def broadcast(self, obj: Any) -> None: if BACKPRESSURE_WARNING and (delta > BACKPRESSURE_REFRACTORY): logger.warning(f"{self.topic} under subscriber backpressure!") self._last_backpressure_event = time.time() + trace_backpressure = self._profile._trace_backpressure_wait_enabled + wait_start_ns = PROFILE_TIME() if trace_backpressure else None await self._backpressure.wait(buf_idx) + if trace_backpressure and wait_start_ns is not None: + self._profile.record_backpressure_wait( + PROFILE_TIME() - wait_start_ns, msg_seq=self._msg_id + ) - # Get local channel and put variable there for local tx - self._local_channel.put_local(self._msg_id, obj) + if self._should_use_local_fast_path(): + self._local_channel.put_local(self._msg_id, obj) - if self._force_tcp or any( - ch.pid != self.pid or not ch.shm_ok for ch in self._channels.values() - ): + if any(not self._can_deliver_locally(ch) for ch in self._channels.values()): with MessageMarshal.serialize(self._msg_id, obj) as ( total_size, header, @@ -449,9 +511,7 @@ async def broadcast(self, obj: Any) -> None: ): total_size_bytes = uint64_to_bytes(total_size) - if not self._force_tcp and any( - ch.pid != self.pid and ch.shm_ok for ch in self._channels.values() - ): + if any(self._can_deliver_via_shm(ch) for ch in self._channels.values()): if self._shm.buf_size < total_size: new_shm = await GraphService(self._graph_address).create_shm( self._num_buffers, total_size * 2 @@ -475,14 +535,10 @@ async def broadcast(self, obj: Any) -> None: for channel in self._channels.values(): msg: bytes = b"" - if self.pid == channel.pid and channel.shm_ok: + if self._can_deliver_locally(channel): continue # Local transmission handled by channel.put - elif ( - (not self._force_tcp) - and self.pid != channel.pid - and channel.shm_ok - ): + elif self._can_deliver_via_shm(channel): msg = ( Command.TX_SHM.value + msg_id_bytes @@ -502,10 +558,25 @@ async def broadcast(self, obj: Any) -> None: channel.writer.write(msg) await channel.writer.drain() self._backpressure.lease(channel.id, buf_idx) + self._profile.sample_inflight(self._backpressure.pressure) except (ConnectionResetError, BrokenPipeError): logger.debug( f"Publisher {self.id}: Channel {channel.id} connection fail" ) + self._profile.record_publish(self._backpressure.pressure, msg_seq=self._msg_id) self._msg_id += 1 + + def _should_use_local_fast_path(self) -> bool: + return any(self._can_deliver_locally(ch) for ch in self._channels.values()) + + def _can_deliver_locally(self, channel: PubChannelInfo) -> bool: + return self._allow_local and self.pid == channel.pid and channel.shm_ok + + def _can_deliver_via_shm(self, channel: PubChannelInfo) -> bool: + return ( + (not self._force_tcp) + and channel.shm_ok + and not self._can_deliver_locally(channel) + ) diff --git a/src/ezmsg/core/relay.py b/src/ezmsg/core/relay.py new file mode 100644 index 00000000..830b7bb6 --- /dev/null +++ b/src/ezmsg/core/relay.py @@ -0,0 +1,42 @@ +from copy import deepcopy +from collections.abc import AsyncGenerator +from typing import Any + +from .settings import Settings +from .netprotocol import DEFAULT_SHM_SIZE +from .stream import InputStream, OutputStream +from .unit import Unit, publisher, subscriber + + +class _RelaySettings(Settings): + leaky: bool = False + max_queue: int | None = None + host: str | None = None + port: int | None = None + num_buffers: int = 32 + buf_size: int = DEFAULT_SHM_SIZE + force_tcp: bool = False + copy_on_forward: bool = True + + +class _CollectionRelayUnit(Unit): + SETTINGS = _RelaySettings + + INPUT = InputStream(Any) + OUTPUT = OutputStream(Any) + + async def initialize(self) -> None: + self.INPUT.leaky = self.SETTINGS.leaky + self.INPUT.max_queue = self.SETTINGS.max_queue + self.OUTPUT.host = self.SETTINGS.host + self.OUTPUT.port = self.SETTINGS.port + self.OUTPUT.num_buffers = self.SETTINGS.num_buffers + self.OUTPUT.buf_size = self.SETTINGS.buf_size + self.OUTPUT.force_tcp = self.SETTINGS.force_tcp + + @subscriber(INPUT) + @publisher(OUTPUT) + async def relay(self, msg: Any) -> AsyncGenerator: + if self.SETTINGS.copy_on_forward: + msg = deepcopy(msg) + yield self.OUTPUT, msg diff --git a/src/ezmsg/core/settingsmeta.py b/src/ezmsg/core/settingsmeta.py new file mode 100644 index 00000000..1260e8ef --- /dev/null +++ b/src/ezmsg/core/settingsmeta.py @@ -0,0 +1,329 @@ +from __future__ import annotations + +from dataclasses import MISSING, asdict, fields as dataclass_fields, is_dataclass +import enum +from collections.abc import Mapping +from typing import Any, get_args, get_origin + +from .graphmeta import SettingsFieldMetadata, SettingsSchemaMetadata + + +def _type_name(tp: object) -> str: + if isinstance(tp, type): + return f"{tp.__module__}.{tp.__qualname__}" + return str(tp) + + +def _sanitize(value: Any) -> Any: + if value is None or isinstance(value, (bool, int, float, str)): + return value + if isinstance(value, enum.Enum): + return _sanitize(value.value) + if isinstance(value, Mapping): + return {str(key): _sanitize(val) for key, val in value.items()} + if isinstance(value, (list, tuple, set, frozenset)): + return [_sanitize(val) for val in value] + if is_dataclass(value): + try: + return _sanitize(asdict(value)) + except Exception: + return repr(value) + return repr(value) + + +def settings_structured_value(value: object) -> dict[str, Any] | None: + if value is None: + return None + + if is_dataclass(value): + try: + asdict_value = asdict(value) + if isinstance(asdict_value, dict): + return _sanitize(asdict_value) + except Exception: + pass + + if hasattr(value, "model_dump") and callable(getattr(value, "model_dump")): + try: + dumped = value.model_dump() # type: ignore[attr-defined] + if isinstance(dumped, dict): + return _sanitize(dumped) + except Exception: + pass + + if hasattr(value, "dict") and callable(getattr(value, "dict")): + try: + dumped = value.dict() # type: ignore[attr-defined] + if isinstance(dumped, dict): + return _sanitize(dumped) + except Exception: + pass + + if isinstance(value, Mapping): + return _sanitize(dict(value)) + + if hasattr(value, "param"): + param_ns = getattr(value, "param") + if hasattr(param_ns, "values") and callable(param_ns.values): + try: + values = param_ns.values() + if isinstance(values, dict): + return _sanitize(values) + except Exception: + pass + + return None + + +def settings_repr_value(value: object) -> dict[str, Any] | str: + structured = settings_structured_value(value) + if structured is not None: + return structured + return repr(value) + + +def _widget_hint( + *, + field_type: str, + choices: list[Any] | None, + bounds: tuple[float | None, float | None] | None, +) -> str | None: + field_type_lower = field_type.lower() + if choices: + return "select" + if "bool" in field_type_lower: + return "checkbox" + if bounds is not None and ("int" in field_type_lower or "float" in field_type_lower): + return "slider" + if "int" in field_type_lower: + return "int_input" + if "float" in field_type_lower: + return "float_input" + if "str" in field_type_lower: + return "text_input" + return None + + +def _choices_from_annotation(annotation: Any) -> list[Any] | None: + origin = get_origin(annotation) + if origin is None: + return None + origin_name = getattr(origin, "__name__", str(origin)) + if origin_name != "Literal": + return None + return [_sanitize(val) for val in get_args(annotation)] + + +def _extract_bounds(obj: object) -> tuple[float | None, float | None] | None: + lower = None + upper = None + for attr in ("ge", "gt", "min_length"): + if hasattr(obj, attr): + bound_val = getattr(obj, attr) + if isinstance(bound_val, (int, float)): + lower = float(bound_val) + break + for attr in ("le", "lt", "max_length"): + if hasattr(obj, attr): + bound_val = getattr(obj, attr) + if isinstance(bound_val, (int, float)): + upper = float(bound_val) + break + if lower is None and upper is None: + return None + return (lower, upper) + + +def settings_schema_from_type(settings_type: object) -> SettingsSchemaMetadata | None: + if not isinstance(settings_type, type): + return None + + if is_dataclass(settings_type): + fields: list[SettingsFieldMetadata] = [] + for f in dataclass_fields(settings_type): + required = f.default is MISSING and f.default_factory is MISSING + default_val: Any | None = None + if not required: + if f.default is not MISSING: + default_val = _sanitize(f.default) + elif f.default_factory is not MISSING: + try: + default_val = _sanitize(f.default_factory()) + except Exception: + default_val = "" + metadata = f.metadata if isinstance(f.metadata, Mapping) else {} + description = metadata.get("description") + choices = metadata.get("choices") + if isinstance(choices, (list, tuple, set)): + choices = [_sanitize(val) for val in choices] + else: + choices = _choices_from_annotation(f.type) + bounds = None + ge = metadata.get("ge", metadata.get("min")) + le = metadata.get("le", metadata.get("max")) + if isinstance(ge, (int, float)) or isinstance(le, (int, float)): + bounds = ( + float(ge) if isinstance(ge, (int, float)) else None, + float(le) if isinstance(le, (int, float)) else None, + ) + field_type = _type_name(f.type) + fields.append( + SettingsFieldMetadata( + name=f.name, + field_type=field_type, + required=required, + default=default_val, + description=description if isinstance(description, str) else None, + bounds=bounds, + choices=choices if isinstance(choices, list) else None, + widget_hint=_widget_hint( + field_type=field_type, + choices=choices if isinstance(choices, list) else None, + bounds=bounds, + ), + ) + ) + return SettingsSchemaMetadata( + provider="dataclass", + settings_type=_type_name(settings_type), + fields=fields, + ) + + if hasattr(settings_type, "model_fields"): + model_fields = getattr(settings_type, "model_fields") + if isinstance(model_fields, dict): + fields: list[SettingsFieldMetadata] = [] + for name, field_info in model_fields.items(): + annotation = getattr(field_info, "annotation", Any) + is_required_attr = getattr(field_info, "is_required", None) + required = ( + bool(is_required_attr()) + if callable(is_required_attr) + else bool(is_required_attr) + ) + default_val = None + if not required: + default = getattr(field_info, "default", None) + default_val = _sanitize(default) + description = getattr(field_info, "description", None) + choices = _choices_from_annotation(annotation) + bounds = _extract_bounds(field_info) + field_type = _type_name(annotation) + fields.append( + SettingsFieldMetadata( + name=name, + field_type=field_type, + required=required, + default=default_val, + description=description if isinstance(description, str) else None, + bounds=bounds, + choices=choices, + widget_hint=_widget_hint( + field_type=field_type, choices=choices, bounds=bounds + ), + ) + ) + return SettingsSchemaMetadata( + provider="pydantic", + settings_type=_type_name(settings_type), + fields=fields, + ) + + if hasattr(settings_type, "__fields__"): + model_fields = getattr(settings_type, "__fields__") + if isinstance(model_fields, dict): + fields: list[SettingsFieldMetadata] = [] + for name, field_info in model_fields.items(): + annotation = getattr(field_info, "outer_type_", Any) + required = bool(getattr(field_info, "required", False)) + default_val = None if required else _sanitize(getattr(field_info, "default", None)) + fi = getattr(field_info, "field_info", None) + description = getattr(fi, "description", None) if fi is not None else None + choices = _choices_from_annotation(annotation) + bounds = _extract_bounds(fi if fi is not None else field_info) + field_type = _type_name(annotation) + fields.append( + SettingsFieldMetadata( + name=name, + field_type=field_type, + required=required, + default=default_val, + description=description if isinstance(description, str) else None, + bounds=bounds, + choices=choices, + widget_hint=_widget_hint( + field_type=field_type, choices=choices, bounds=bounds + ), + ) + ) + return SettingsSchemaMetadata( + provider="pydantic", + settings_type=_type_name(settings_type), + fields=fields, + ) + + param_ns = getattr(settings_type, "param", None) + if param_ns is not None and hasattr(param_ns, "objects"): + try: + objects = param_ns.objects("existing") + except Exception: + try: + objects = param_ns.objects() + except Exception: + objects = None + if isinstance(objects, dict): + fields: list[SettingsFieldMetadata] = [] + for name, param_obj in objects.items(): + if name == "name": + continue + choices_obj = getattr(param_obj, "objects", None) + choices = None + if isinstance(choices_obj, Mapping): + choices = [_sanitize(choice) for choice in choices_obj.keys()] + elif isinstance(choices_obj, (list, tuple, set)): + choices = [_sanitize(choice) for choice in choices_obj] + bounds_obj = getattr(param_obj, "bounds", None) + bounds = None + if ( + isinstance(bounds_obj, tuple) + and len(bounds_obj) == 2 + and all( + bound is None or isinstance(bound, (int, float)) + for bound in bounds_obj + ) + ): + bounds = ( + float(bounds_obj[0]) if isinstance(bounds_obj[0], (int, float)) else None, + float(bounds_obj[1]) if isinstance(bounds_obj[1], (int, float)) else None, + ) + default_val = _sanitize(getattr(param_obj, "default", None)) + description = getattr(param_obj, "doc", None) + field_type = _type_name(type(param_obj)) + fields.append( + SettingsFieldMetadata( + name=name, + field_type=field_type, + required=False, + default=default_val, + description=description if isinstance(description, str) else None, + bounds=bounds, + choices=choices, + widget_hint=_widget_hint( + field_type=field_type, choices=choices, bounds=bounds + ), + ) + ) + return SettingsSchemaMetadata( + provider="param", + settings_type=_type_name(settings_type), + fields=fields, + ) + + return None + + +def settings_schema_from_value(value: object) -> SettingsSchemaMetadata | None: + if value is None: + return None + return settings_schema_from_type(type(value)) + diff --git a/src/ezmsg/core/stream.py b/src/ezmsg/core/stream.py index c719c92a..edce08c2 100644 --- a/src/ezmsg/core/stream.py +++ b/src/ezmsg/core/stream.py @@ -26,6 +26,110 @@ def __repr__(self) -> str: return f"Stream:{_addr}[{self.msg_type.__name__}]" +class Topic(Stream): + """ + Graph endpoint metadata for Collection boundaries and graph wiring. + + Topics represent named DAG nodes only. Unlike InputStream / OutputStream, + they do not directly configure Subscriber / Publisher transport behavior. + """ + + def __repr__(self) -> str: + return f"Topic{super().__repr__()}()" + + +class InputTopic(Topic): + """ + Directional alias for a Collection input topic. + """ + + def __repr__(self) -> str: + return f"Input{super().__repr__()}" + + +class OutputTopic(Topic): + """ + Directional alias for a Collection output topic. + """ + + def __repr__(self) -> str: + return f"Output{super().__repr__()}" + + +class InputRelay(InputTopic): + """ + Collection input boundary that materializes an internal relay subscriber/publisher. + + This enables subscriber-side behavior (e.g., leaky reception) on the boundary. + """ + + leaky: bool + max_queue: int | None + copy_on_forward: bool + + def __init__( + self, + msg_type: Any, + leaky: bool = False, + max_queue: int | None = None, + copy_on_forward: bool = True, + ) -> None: + super().__init__(msg_type) + if max_queue is not None and max_queue <= 0: + raise ValueError("max_queue must be positive") + self.leaky = leaky + self.max_queue = max_queue + self.copy_on_forward = copy_on_forward + + def __repr__(self) -> str: + base = f"InputRelay{Stream.__repr__(self)}" + return ( + f"{base}(leaky={self.leaky}, max_queue={self.max_queue}, " + f"copy_on_forward={self.copy_on_forward})" + ) + + +class OutputRelay(OutputTopic): + """ + Collection output boundary that materializes an internal relay subscriber/publisher. + + This enables publisher-side behavior (e.g., custom transport buffer settings) + on the boundary. + """ + + host: str | None + port: int | None + num_buffers: int + buf_size: int + force_tcp: bool + copy_on_forward: bool + + def __init__( + self, + msg_type: Any, + host: str | None = None, + port: int | None = None, + num_buffers: int = 32, + buf_size: int = DEFAULT_SHM_SIZE, + force_tcp: bool = False, + copy_on_forward: bool = True, + ) -> None: + super().__init__(msg_type) + self.host = host + self.port = port + self.num_buffers = num_buffers + self.buf_size = buf_size + self.force_tcp = force_tcp + self.copy_on_forward = copy_on_forward + + def __repr__(self) -> str: + base = f"OutputRelay{Stream.__repr__(self)}" + return ( + f"{base}(num_buffers={self.num_buffers}, force_tcp={self.force_tcp}, " + f"copy_on_forward={self.copy_on_forward})" + ) + + class InputStream(Stream): """ Can be added to any Component as a member variable. Methods may subscribe to it. @@ -124,15 +228,20 @@ class OutputStream(Stream): :type num_buffers: int :param buf_size: Size of each message buffer in bytes :type buf_size: int - :param force_tcp: Whether to force TCP transport instead of shared memory - :type force_tcp: bool + :param force_tcp: Whether to force TCP transport instead of shared memory. + If None, inherit the process default from ``EZMSG_FORCE_TCP``. + :type force_tcp: bool | None + :param allow_local: Whether to allow the in-process fast path when available. + If None, inherit the process default from ``EZMSG_ALLOW_LOCAL``. + :type allow_local: bool | None """ host: str | None port: int | None num_buffers: int buf_size: int - force_tcp: bool + force_tcp: bool | None + allow_local: bool | None def __init__( self, @@ -141,7 +250,8 @@ def __init__( port: int | None = None, num_buffers: int = 32, buf_size: int = DEFAULT_SHM_SIZE, - force_tcp: bool = False, + force_tcp: bool | None = None, + allow_local: bool | None = None, ) -> None: super().__init__(msg_type) self.host = host @@ -149,7 +259,11 @@ def __init__( self.num_buffers = num_buffers self.buf_size = buf_size self.force_tcp = force_tcp + self.allow_local = allow_local def __repr__(self) -> str: preamble = f"Output{super().__repr__()}" - return f"{preamble}({self.num_buffers=}, {self.force_tcp=})" + return ( + f"{preamble}({self.num_buffers=}, {self.force_tcp=}, " + f"{self.allow_local=})" + ) diff --git a/src/ezmsg/core/subclient.py b/src/ezmsg/core/subclient.py index 3ca2dc22..a3ad3246 100644 --- a/src/ezmsg/core/subclient.py +++ b/src/ezmsg/core/subclient.py @@ -9,6 +9,7 @@ from .graphserver import GraphService from .channelmanager import CHANNELS from .messagechannel import NotificationQueue, LeakyQueue, Channel +from .profiling import PROFILES, PROFILE_TIME from .netprotocol import ( AddressType, @@ -40,6 +41,7 @@ class Subscriber: _graph_address: AddressType | None _graph_task: asyncio.Task[None] _incoming: NotificationQueue + _profile: object # FIXME: This event allows Subscriber.create to block until # incoming initial connections (UPDATE) has completed. The @@ -128,6 +130,8 @@ def __init__( self._graph_address = graph_address self._channels = dict() + self._active_msg_seq: int | None = None + self._active_trace_sampled = False if self.leaky: self._incoming = LeakyQueue( 1 if max_queue is None else max_queue, self._handle_dropped_notification @@ -135,6 +139,7 @@ def __init__( else: self._incoming = asyncio.Queue() self._initialized = asyncio.Event() + self._profile = PROFILES.register_subscriber(self.id, self.topic) def _handle_dropped_notification( self, notification: typing.Tuple[UUID, int] @@ -160,6 +165,7 @@ def close(self) -> None: and closes all shared memory contexts. """ self._graph_task.cancel() + PROFILES.unregister_subscriber(self.id) async def wait_closed(self) -> None: """ @@ -295,5 +301,40 @@ async def recv_zero_copy(self) -> typing.AsyncGenerator[typing.Any, None]: break # Stale notification from an unregistered publisher — skip. - with self._channels[pub_id].get(msg_id, self.id) as msg: - yield msg + channel = self._channels[pub_id] + channel_kind = channel.channel_kind + self._active_msg_seq = msg_id + self._active_trace_sampled = self._profile.begin_message(channel_kind) + try: + trace_lease = self._profile._trace_lease_time_enabled + start_ns = PROFILE_TIME() if trace_lease else None + with channel.get(msg_id, self.id) as msg: + yield msg + lease_ns = None + if trace_lease and start_ns is not None: + lease_ns = PROFILE_TIME() - start_ns + self._profile.record_lease_time( + channel_kind, + lease_ns, + msg_seq=msg_id, + sampled=self._active_trace_sampled, + ) + finally: + self._active_msg_seq = None + self._active_trace_sampled = False + + def begin_profile(self) -> int: + if not self._profile._trace_user_span_enabled or not self._active_trace_sampled: + return 0 + return PROFILE_TIME() + + def end_profile(self, start_ns: int, label: str | None = None) -> None: + if start_ns <= 0: + return + end_ns = PROFILE_TIME() + self._profile.record_user_span( + end_ns - start_ns, + label, + msg_seq=self._active_msg_seq, + sampled=self._active_trace_sampled, + ) diff --git a/src/ezmsg/core/unit.py b/src/ezmsg/core/unit.py index 8527e71d..0fc8b66b 100644 --- a/src/ezmsg/core/unit.py +++ b/src/ezmsg/core/unit.py @@ -2,7 +2,7 @@ import inspect import functools import warnings -from .stream import InputStream, OutputStream +from .stream import InputStream, OutputStream, Topic from .component import ComponentMeta, Component from .settings import Settings @@ -55,6 +55,12 @@ def __init__( cls.__threads__[thread_name] = thread for field_name, field_value in fields.items(): + if isinstance(field_value, Topic): + raise TypeError( + f"{name}.{field_name} is a {type(field_value).__name__}. " + "Units may only declare InputStream / OutputStream endpoints. " + "Use Topic / Relay endpoints on Collections only." + ) if callable(field_value): if hasattr(field_value, TASK_ATTR): cls.__tasks__[field_name] = field_value diff --git a/src/ezmsg/util/perf/ab.py b/src/ezmsg/util/perf/ab.py new file mode 100644 index 00000000..5a2da621 --- /dev/null +++ b/src/ezmsg/util/perf/ab.py @@ -0,0 +1,446 @@ +from __future__ import annotations + +import argparse +import contextlib +import json +import os +import random +import shutil +import subprocess +import sys +import tempfile + +from dataclasses import asdict, dataclass +from pathlib import Path + + +DEFAULT_PAIR_SEED = 0 + + +@dataclass(frozen=True) +class ABCaseSummary: + case_id: str + a_us_per_message_median: float + b_us_per_message_median: float + delta_pct_median: float + delta_pct_mean: float + pair_count: int + b_faster_pairs: int + + +@dataclass(frozen=True) +class ABRunSummary: + ref_a: str + ref_b: str + rounds: int + seed: int + cases: list[ABCaseSummary] + + +def build_pair_order(rounds: int, seed: int) -> list[tuple[str, str]]: + base = [("A", "B"), ("B", "A")] * ((rounds + 1) // 2) + order = base[:rounds] + random.Random(seed).shuffle(order) + return order + + +def _hotpath_json_arg(path: Path) -> list[str]: + return ["--json-out", str(path)] + + +def build_hotpath_command( + output_path: Path, + count: int, + warmup: int, + payload_sizes: list[int], + transports: list[str], + apis: list[str], + num_buffers: int, + quiet: bool, +) -> list[str]: + cmd = [ + "uv", + "run", + "python", + "-m", + "ezmsg.util.perf.hotpath", + "--count", + str(count), + "--warmup", + str(warmup), + "--samples", + "1", + "--num-buffers", + str(num_buffers), + "--payload-sizes", + *[str(payload_size) for payload_size in payload_sizes], + "--transports", + *transports, + "--apis", + *apis, + *_hotpath_json_arg(output_path), + ] + if quiet: + cmd.append("--quiet") + return cmd + + +def load_hotpath_summary(path: Path) -> dict[str, float]: + payload = json.loads(path.read_text()) + return { + entry["case_id"]: float(entry["summary"]["us_per_message_median"]) + for entry in payload["results"] + } + + +def summarize_ab_results( + ref_a: str, + ref_b: str, + rounds: int, + seed: int, + paired_runs: list[tuple[dict[str, float], dict[str, float]]], +) -> ABRunSummary: + case_ids = sorted(paired_runs[0][0].keys()) + cases: list[ABCaseSummary] = [] + + for case_id in case_ids: + a_values = [pair[0][case_id] for pair in paired_runs] + b_values = [pair[1][case_id] for pair in paired_runs] + deltas = [((b / a) - 1.0) * 100.0 for a, b in zip(a_values, b_values)] + cases.append( + ABCaseSummary( + case_id=case_id, + a_us_per_message_median=_median(a_values), + b_us_per_message_median=_median(b_values), + delta_pct_median=_median(deltas), + delta_pct_mean=sum(deltas) / len(deltas), + pair_count=len(deltas), + b_faster_pairs=sum(1 for a, b in zip(a_values, b_values) if b < a), + ) + ) + + return ABRunSummary( + ref_a=ref_a, + ref_b=ref_b, + rounds=rounds, + seed=seed, + cases=cases, + ) + + +def _median(values: list[float]) -> float: + ordered = sorted(values) + mid = len(ordered) // 2 + if len(ordered) % 2: + return ordered[mid] + return (ordered[mid - 1] + ordered[mid]) / 2.0 + + +def _run_checked(cmd: list[str], cwd: Path) -> None: + env = os.environ.copy() + env.pop("VIRTUAL_ENV", None) + completed = subprocess.run(cmd, cwd=cwd, capture_output=True, text=True, env=env) + if completed.returncode == 0: + return + + raise RuntimeError( + f"Command failed in {cwd}:\n" + f"$ {' '.join(cmd)}\n\n" + f"stdout:\n{completed.stdout}\n" + f"stderr:\n{completed.stderr}" + ) + + +def _is_current_ref(ref: str) -> bool: + return ref.upper() == "CURRENT" + + +@contextlib.contextmanager +def _provision_tree( + repo_root: Path, + ref: str, + label: str, + keep: bool, +) -> Path: + if _is_current_ref(ref): + yield repo_root + return + + parent = Path(tempfile.mkdtemp(prefix=f"ezmsg-perf-{label.lower()}-")) + tree_path = parent / "tree" + _run_checked( + ["git", "worktree", "add", "--detach", str(tree_path), ref], + cwd=repo_root, + ) + + try: + yield tree_path + finally: + if keep: + return + try: + _run_checked(["git", "worktree", "remove", "--force", str(tree_path)], cwd=repo_root) + finally: + shutil.rmtree(parent, ignore_errors=True) + + +def _maybe_sync(tree: Path) -> None: + _run_checked(["uv", "sync", "--group", "dev"], cwd=tree) + + +def _mirror_hotpath_module(source_root: Path, target_tree: Path) -> None: + source = source_root / "src" / "ezmsg" / "util" / "perf" / "hotpath.py" + target = target_tree / "src" / "ezmsg" / "util" / "perf" / "hotpath.py" + shutil.copy2(source, target) + + +def _ensure_json_files_match( + left: dict[str, float], + right: dict[str, float], + label_left: str, + label_right: str, +) -> None: + if left.keys() == right.keys(): + return + + raise RuntimeError( + f"Benchmark cases differ between {label_left} and {label_right}: " + f"{sorted(left.keys())} != {sorted(right.keys())}" + ) + + +def _print_summary(summary: ABRunSummary) -> None: + print( + f"Interleaved hot-path comparison: A={summary.ref_a}, " + f"B={summary.ref_b}, rounds={summary.rounds}, seed={summary.seed}" + ) + for case in summary.cases: + sign = "regression" if case.delta_pct_median > 0 else "improvement" + print( + f"{case.case_id:<36} " + f"A={case.a_us_per_message_median:>10.2f} us/msg " + f"B={case.b_us_per_message_median:>10.2f} us/msg " + f"delta={case.delta_pct_median:>7.2f}% ({sign}) " + f"wins={case.b_faster_pairs}/{case.pair_count}" + ) + + +def dump_ab_json(summary: ABRunSummary, path: Path) -> None: + payload = { + "suite": "hotpath-ab", + "ref_a": summary.ref_a, + "ref_b": summary.ref_b, + "rounds": summary.rounds, + "seed": summary.seed, + "cases": [asdict(case) for case in summary.cases], + } + path.write_text(json.dumps(payload, indent=2) + "\n") + + +def perf_ab( + ref_a: str, + ref_b: str, + rounds: int, + count: int, + warmup: int, + prewarm: int, + payload_sizes: list[int], + transports: list[str], + apis: list[str], + num_buffers: int, + seed: int, + json_out: Path | None, + keep_worktrees: bool, + sync: bool, + quiet: bool, +) -> None: + if rounds <= 0: + raise ValueError("rounds must be > 0") + + repo_root = Path( + subprocess.run( + ["git", "rev-parse", "--show-toplevel"], + check=True, + capture_output=True, + text=True, + ).stdout.strip() + ) + pair_order = build_pair_order(rounds, seed) + + with _provision_tree(repo_root, ref_a, "A", keep_worktrees) as tree_a: + with _provision_tree(repo_root, ref_b, "B", keep_worktrees) as tree_b: + if tree_a != repo_root: + _mirror_hotpath_module(repo_root, tree_a) + if tree_b != repo_root: + _mirror_hotpath_module(repo_root, tree_b) + + if sync: + _maybe_sync(tree_a) + if tree_b != tree_a: + _maybe_sync(tree_b) + + with tempfile.TemporaryDirectory(prefix="ezmsg-perf-ab-runs-") as tmpdir_name: + tmpdir = Path(tmpdir_name) + cmd_by_label = { + "A": lambda path: build_hotpath_command( + path, + count=count, + warmup=warmup, + payload_sizes=payload_sizes, + transports=transports, + apis=apis, + num_buffers=num_buffers, + quiet=quiet, + ), + "B": lambda path: build_hotpath_command( + path, + count=count, + warmup=warmup, + payload_sizes=payload_sizes, + transports=transports, + apis=apis, + num_buffers=num_buffers, + quiet=quiet, + ), + } + tree_by_label = {"A": tree_a, "B": tree_b} + + for idx in range(prewarm): + for label in ("A", "B"): + if label == "B" and tree_b == tree_a: + continue + warm_path = tmpdir / f"warm-{label}-{idx}.json" + _run_checked(cmd_by_label[label](warm_path), cwd=tree_by_label[label]) + + paired_runs: list[tuple[dict[str, float], dict[str, float]]] = [] + for round_idx, (first, second) in enumerate(pair_order, start=1): + outputs: dict[str, dict[str, float]] = {} + for label in (first, second): + output_path = tmpdir / f"round-{round_idx:02d}-{label}.json" + _run_checked(cmd_by_label[label](output_path), cwd=tree_by_label[label]) + outputs[label] = load_hotpath_summary(output_path) + + _ensure_json_files_match(outputs["A"], outputs["B"], ref_a, ref_b) + paired_runs.append((outputs["A"], outputs["B"])) + + summary = summarize_ab_results(ref_a, ref_b, rounds, seed, paired_runs) + _print_summary(summary) + if json_out is not None: + dump_ab_json(summary, json_out) + print(f"Wrote JSON results to {json_out}") + + +def setup_ab_cmdline(subparsers: argparse._SubParsersAction) -> None: + p_ab = subparsers.add_parser( + "ab", + help="run interleaved A/B hot-path comparisons using git worktrees", + ) + p_ab.add_argument("--ref-a", default="dev", help="baseline git ref or CURRENT") + p_ab.add_argument("--ref-b", default="CURRENT", help="candidate git ref or CURRENT") + p_ab.add_argument( + "--rounds", + type=int, + default=6, + help="number of A/B pairs to run (default = 6)", + ) + p_ab.add_argument( + "--count", + type=int, + default=2_000, + help="messages per hot-path sample (default = 2000)", + ) + p_ab.add_argument( + "--warmup", + type=int, + default=200, + help="warmup messages per hot-path sample (default = 200)", + ) + p_ab.add_argument( + "--prewarm", + type=int, + default=1, + help="unmeasured warmup invocations per side (default = 1)", + ) + p_ab.add_argument( + "--payload-sizes", + nargs="*", + type=int, + default=[64, 4096], + help="payload sizes in bytes (default = [64, 4096])", + ) + p_ab.add_argument( + "--transports", + nargs="*", + choices=["local", "shm", "tcp"], + default=["local", "shm", "tcp"], + help="transports to compare (default = ['local', 'shm', 'tcp'])", + ) + p_ab.add_argument( + "--apis", + nargs="*", + choices=["async", "sync"], + default=["async"], + help="apis to compare (default = ['async'])", + ) + p_ab.add_argument( + "--num-buffers", + type=int, + default=1, + help="publisher buffers (default = 1)", + ) + p_ab.add_argument( + "--seed", + type=int, + default=DEFAULT_PAIR_SEED, + help="pair-order shuffle seed (default = 0)", + ) + p_ab.add_argument( + "--json-out", + type=Path, + default=None, + help="optional JSON output path", + ) + p_ab.add_argument( + "--keep-worktrees", + action="store_true", + help="leave auto-provisioned worktrees on disk for inspection", + ) + p_ab.add_argument( + "--sync", + action="store_true", + help="run 'uv sync --group dev' in each provisioned worktree first", + ) + p_ab.add_argument( + "--quiet", + action="store_true", + help="suppress ezmsg runtime logs in child benchmark runs", + ) + p_ab.set_defaults( + _handler=lambda ns: perf_ab( + ref_a=ns.ref_a, + ref_b=ns.ref_b, + rounds=ns.rounds, + count=ns.count, + warmup=ns.warmup, + prewarm=ns.prewarm, + payload_sizes=ns.payload_sizes, + transports=ns.transports, + apis=ns.apis, + num_buffers=ns.num_buffers, + seed=ns.seed, + json_out=ns.json_out, + keep_worktrees=ns.keep_worktrees, + sync=ns.sync, + quiet=ns.quiet, + ) + ) + + +def main() -> None: + parser = argparse.ArgumentParser( + description="Run interleaved ezmsg hot-path A/B comparisons." + ) + subparsers = parser.add_subparsers(dest="command", required=True) + setup_ab_cmdline(subparsers) + ns = parser.parse_args(["ab", *sys.argv[1:]]) + ns._handler(ns) diff --git a/src/ezmsg/util/perf/command.py b/src/ezmsg/util/perf/command.py index 21fed7eb..9dab1f8e 100644 --- a/src/ezmsg/util/perf/command.py +++ b/src/ezmsg/util/perf/command.py @@ -1,6 +1,8 @@ import argparse +from .ab import setup_ab_cmdline from .analysis import setup_summary_cmdline +from .hotpath import setup_hotpath_cmdline from .run import setup_run_cmdline @@ -9,6 +11,8 @@ def command() -> None: subparsers = parser.add_subparsers(dest="command", required=True) setup_run_cmdline(subparsers) + setup_hotpath_cmdline(subparsers) + setup_ab_cmdline(subparsers) setup_summary_cmdline(subparsers) ns = parser.parse_args() diff --git a/src/ezmsg/util/perf/hotpath.py b/src/ezmsg/util/perf/hotpath.py new file mode 100644 index 00000000..74585981 --- /dev/null +++ b/src/ezmsg/util/perf/hotpath.py @@ -0,0 +1,437 @@ +from __future__ import annotations + +import argparse +import asyncio +import contextlib +import inspect +import json +import logging +import random +import statistics +import sys +import time + +from dataclasses import asdict, dataclass +from pathlib import Path +from typing import Literal +from uuid import uuid4 + +import ezmsg.core as ez + +from ezmsg.core.graphserver import GraphServer + +from .util import coef_var, median_of_means, stable_perf + +ApiName = Literal["async", "sync"] +TransportName = Literal["local", "shm", "tcp"] + +DEFAULT_APIS: tuple[ApiName, ...] = ("async",) +DEFAULT_TRANSPORTS: tuple[TransportName, ...] = ("local", "shm", "tcp") +DEFAULT_PAYLOAD_SIZES = (64, 4096) + + +def _supports_same_process_transport_selection() -> bool: + return "allow_local" in inspect.signature(ez.Publisher.create).parameters + + +def _validate_transport_support(case: "HotPathCase") -> None: + if case.transport == "local": + return + if _supports_same_process_transport_selection(): + return + raise RuntimeError( + "This ref does not support bypassing the same-process local fast path, " + f"so '{case.case_id}' cannot be benchmarked here. Compare against a ref " + "that includes the allow_local transport-selection support, or limit the " + "run to '--transports local'." + ) + + +def _publisher_transport_kwargs(case: "HotPathCase", host: str) -> dict[str, object]: + kwargs: dict[str, object] = { + "host": host, + "num_buffers": case.num_buffers, + "force_tcp": (case.transport == "tcp"), + } + if _supports_same_process_transport_selection(): + kwargs["allow_local"] = case.transport == "local" + return kwargs + + +@dataclass(frozen=True) +class HotPathCase: + api: ApiName + transport: TransportName + payload_size: int + num_buffers: int + + @property + def case_id(self) -> str: + return ( + f"{self.api}/{self.transport}/payload={self.payload_size}" + f"/buffers={self.num_buffers}" + ) + + +@dataclass(frozen=True) +class HotPathSummary: + samples: int + seconds_median: float + seconds_mean: float + seconds_min: float + seconds_max: float + seconds_median_of_means: float + seconds_cv: float + us_per_message_median: float + us_per_message_mean: float + messages_per_second_median: float + messages_per_second_mean: float + + +@dataclass(frozen=True) +class HotPathCaseResult: + case: HotPathCase + count: int + warmup: int + samples_seconds: list[float] + summary: HotPathSummary + + +@dataclass(frozen=True) +class HotPathSuiteResult: + seed: int + count: int + warmup: int + results: list[HotPathCaseResult] + + +@contextlib.contextmanager +def _quiet_ezmsg_logs(enabled: bool): + if not enabled: + yield + return + + old_level = ez.logger.level + ez.logger.setLevel(logging.WARNING) + try: + yield + finally: + ez.logger.setLevel(old_level) + + +def build_cases( + apis: list[str], + transports: list[str], + payload_sizes: list[int], + num_buffers: int, +) -> list[HotPathCase]: + cases = [ + HotPathCase( + api=api, # type: ignore[arg-type] + transport=transport, # type: ignore[arg-type] + payload_size=payload_size, + num_buffers=num_buffers, + ) + for api in apis + for transport in transports + for payload_size in payload_sizes + if api in ("async", "sync") and transport in ("local", "shm", "tcp") + ] + return sorted(cases, key=lambda case: case.case_id) + + +def summarize_samples(samples: list[float], count: int) -> HotPathSummary: + us_per_message = [(sample / count) * 1e6 for sample in samples] + rates = [count / sample for sample in samples] + return HotPathSummary( + samples=len(samples), + seconds_median=statistics.median(samples), + seconds_mean=statistics.fmean(samples), + seconds_min=min(samples), + seconds_max=max(samples), + seconds_median_of_means=median_of_means(samples), + seconds_cv=coef_var(samples), + us_per_message_median=statistics.median(us_per_message), + us_per_message_mean=statistics.fmean(us_per_message), + messages_per_second_median=statistics.median(rates), + messages_per_second_mean=statistics.fmean(rates), + ) + + +async def _async_roundtrip( + case: HotPathCase, + count: int, + warmup: int, + graph_address: tuple[str, int], +) -> float: + _validate_transport_support(case) + topic = f"/EZMSG/PERF/HOTPATH/{uuid4().hex}" + payload = bytes(case.payload_size) + + async with ez.GraphContext(graph_address, auto_start=False) as ctx: + pub = await ctx.publisher(topic, **_publisher_transport_kwargs(case, graph_address[0])) + sub = await ctx.subscriber(topic) + + for _ in range(warmup): + await pub.broadcast(payload) + await sub.recv() + + start = time.perf_counter() + for _ in range(count): + await pub.broadcast(payload) + await sub.recv() + return time.perf_counter() - start + + +def _sync_roundtrip( + case: HotPathCase, + count: int, + warmup: int, + graph_address: tuple[str, int], +) -> float: + _validate_transport_support(case) + topic = f"/EZMSG/PERF/HOTPATH/{uuid4().hex}" + payload = bytes(case.payload_size) + + with ez.sync.init(graph_address, auto_start=False) as ctx: + pub = ctx.create_publisher(topic, **_publisher_transport_kwargs(case, graph_address[0])) + sub = ctx.create_subscription(topic) + + for _ in range(warmup): + pub.publish(payload) + sub.recv() + + start = time.perf_counter() + for _ in range(count): + pub.publish(payload) + sub.recv() + return time.perf_counter() - start + + +def run_hotpath_case( + case: HotPathCase, + count: int, + warmup: int, + samples: int, + graph_address: tuple[str, int], +) -> HotPathCaseResult: + results: list[float] = [] + + for _ in range(samples): + with stable_perf(): + if case.api == "async": + elapsed = asyncio.run( + _async_roundtrip(case, count, warmup, graph_address) + ) + else: + elapsed = _sync_roundtrip(case, count, warmup, graph_address) + results.append(elapsed) + + return HotPathCaseResult( + case=case, + count=count, + warmup=warmup, + samples_seconds=results, + summary=summarize_samples(results, count), + ) + + +def _format_case_result(result: HotPathCaseResult) -> str: + summary = result.summary + return ( + f"{result.case.case_id:<36} " + f"{summary.us_per_message_median:>10.2f} us/msg " + f"{summary.messages_per_second_median:>12,.0f} msg/s " + f"cv={summary.seconds_cv:>5.3f}" + ) + + +def run_hotpath_suite( + count: int, + warmup: int, + samples: int, + apis: list[str], + transports: list[str], + payload_sizes: list[int], + num_buffers: int, + seed: int, + quiet: bool, +) -> HotPathSuiteResult: + rng = random.Random(seed) + cases = build_cases(apis, transports, payload_sizes, num_buffers) + rng.shuffle(cases) + + graph_server = GraphServer() + graph_server.start(("127.0.0.1", 0)) + try: + with _quiet_ezmsg_logs(quiet): + results = [ + run_hotpath_case( + case, + count=count, + warmup=warmup, + samples=samples, + graph_address=graph_server.address, + ) + for case in cases + ] + finally: + graph_server.stop() + + results.sort(key=lambda result: result.case.case_id) + return HotPathSuiteResult(seed=seed, count=count, warmup=warmup, results=results) + + +def dump_suite_json(result: HotPathSuiteResult, path: Path) -> None: + payload = { + "suite": "hotpath", + "seed": result.seed, + "count": result.count, + "warmup": result.warmup, + "results": [ + { + "case": asdict(case_result.case), + "case_id": case_result.case.case_id, + "count": case_result.count, + "warmup": case_result.warmup, + "samples_seconds": case_result.samples_seconds, + "summary": asdict(case_result.summary), + } + for case_result in result.results + ], + } + path.write_text(json.dumps(payload, indent=2) + "\n") + + +def perf_hotpath( + count: int, + warmup: int, + samples: int, + apis: list[str], + transports: list[str], + payload_sizes: list[int], + num_buffers: int, + seed: int, + json_out: Path | None, + quiet: bool, +) -> None: + result = run_hotpath_suite( + count=count, + warmup=warmup, + samples=samples, + apis=apis, + transports=transports, + payload_sizes=payload_sizes, + num_buffers=num_buffers, + seed=seed, + quiet=quiet, + ) + + print("Hot-path roundtrip benchmark") + print( + f"count={count}, warmup={warmup}, samples={samples}, " + f"payload_sizes={payload_sizes}, transports={transports}, apis={apis}" + ) + for case_result in result.results: + print(_format_case_result(case_result)) + + if json_out is not None: + dump_suite_json(result, json_out) + print(f"Wrote JSON results to {json_out}") + + +def setup_hotpath_cmdline(subparsers: argparse._SubParsersAction) -> None: + p_hotpath = subparsers.add_parser( + "hotpath", + help="run fast, focused hot-path roundtrip benchmarks", + ) + p_hotpath.add_argument( + "--count", + type=int, + default=5_000, + help="messages per sample (default = 5000)", + ) + p_hotpath.add_argument( + "--warmup", + type=int, + default=500, + help="warmup messages per sample (default = 500)", + ) + p_hotpath.add_argument( + "--samples", + type=int, + default=5, + help="timed samples per case (default = 5)", + ) + p_hotpath.add_argument( + "--apis", + nargs="*", + choices=DEFAULT_APIS + ("sync",), + default=list(DEFAULT_APIS), + help=f"apis to benchmark (default = {list(DEFAULT_APIS)})", + ) + p_hotpath.add_argument( + "--transports", + nargs="*", + choices=DEFAULT_TRANSPORTS, + default=list(DEFAULT_TRANSPORTS), + help=f"transports to benchmark (default = {list(DEFAULT_TRANSPORTS)})", + ) + p_hotpath.add_argument( + "--payload-sizes", + nargs="*", + type=int, + default=list(DEFAULT_PAYLOAD_SIZES), + help=f"payload sizes in bytes (default = {list(DEFAULT_PAYLOAD_SIZES)})", + ) + p_hotpath.add_argument( + "--num-buffers", + type=int, + default=1, + help="publisher buffers (default = 1)", + ) + p_hotpath.add_argument( + "--seed", + type=int, + default=0, + help="shuffle seed for case order (default = 0)", + ) + p_hotpath.add_argument( + "--json-out", + type=Path, + default=None, + help="optional JSON output path", + ) + p_hotpath.add_argument( + "--quiet", + action="store_true", + help="suppress ezmsg runtime logs during the benchmark", + ) + p_hotpath.set_defaults( + _handler=lambda ns: perf_hotpath( + count=ns.count, + warmup=ns.warmup, + samples=ns.samples, + apis=ns.apis, + transports=ns.transports, + payload_sizes=ns.payload_sizes, + num_buffers=ns.num_buffers, + seed=ns.seed, + json_out=ns.json_out, + quiet=ns.quiet, + ) + ) + + +def main() -> None: + parser = argparse.ArgumentParser( + description="Run ezmsg hot-path roundtrip benchmarks." + ) + subparsers = parser.add_subparsers(dest="command", required=True) + setup_hotpath_cmdline(subparsers) + ns = parser.parse_args(["hotpath", *sys.argv[1:]]) + ns._handler(ns) + + +if __name__ == "__main__": + main() diff --git a/src/ezmsg/util/perf/impl.py b/src/ezmsg/util/perf/impl.py index 6bc83d1d..55e2c30b 100644 --- a/src/ezmsg/util/perf/impl.py +++ b/src/ezmsg/util/perf/impl.py @@ -209,7 +209,8 @@ def fanin(config: ConfigSettings) -> Configuration: """many pubs to one sub""" connections: ez.NetworkDefinition = [(config.source.OUTPUT, config.sink.INPUT)] pubs = [LoadTestSource(config.settings) for _ in range(config.n_clients)] - expected_num_msgs = config.sink.SETTINGS.num_msgs * len(pubs) + total_publishers = 1 + len(pubs) + expected_num_msgs = config.sink.SETTINGS.num_msgs * total_publishers config.sink.SETTINGS = replace(config.sink.SETTINGS, num_msgs=expected_num_msgs) # type: ignore for pub in pubs: connections.append((pub.OUTPUT, config.sink.INPUT)) diff --git a/tests/shutdown_runner.py b/tests/shutdown_runner.py index c185a23a..39bb5124 100644 --- a/tests/shutdown_runner.py +++ b/tests/shutdown_runner.py @@ -8,11 +8,14 @@ import ezmsg.core as ez +STARTED = threading.Event() + class BlockingDiskIO(ez.Unit): @ez.task async def blocked_read(self) -> None: # Cross-platform "hung disk I/O" simulation. + STARTED.set() event = threading.Event() self._event = event await asyncio.shield(asyncio.to_thread(event.wait)) @@ -21,6 +24,7 @@ async def blocked_read(self) -> None: class BlockingSocket(ez.Unit): @ez.task async def blocked_recv(self) -> None: + STARTED.set() sock_r, sock_w = socket.socketpair() sock_r.setblocking(True) sock_w.setblocking(True) @@ -33,6 +37,7 @@ async def blocked_recv(self) -> None: class ExplodeOnCancel(ez.Unit): @ez.task async def explode(self) -> None: + STARTED.set() try: while True: await asyncio.sleep(1.0) @@ -43,6 +48,7 @@ async def explode(self) -> None: class StubbornTask(ez.Unit): @ez.task async def ignore_cancel(self) -> None: + STARTED.set() while True: try: await asyncio.sleep(1.0) @@ -84,7 +90,7 @@ def _emit_ready() -> None: def _watch_ready() -> None: while not done.is_set(): - if runner.running: + if runner.running and STARTED.is_set(): _emit_ready() return time.sleep(0.01) diff --git a/tests/test_graph_session.py b/tests/test_graph_session.py new file mode 100644 index 00000000..f3969d32 --- /dev/null +++ b/tests/test_graph_session.py @@ -0,0 +1,106 @@ +import asyncio + +import pytest + +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import GraphMetadata, GraphSnapshot +from ezmsg.core.graphserver import GraphService + + +def _edge_exists(snapshot: GraphSnapshot, from_topic: str, to_topic: str) -> bool: + return to_topic in snapshot.graph.get(from_topic, []) + + +@pytest.mark.asyncio +async def test_session_drop_cleans_owned_edges(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + + await owner.__aenter__() + await observer.__aenter__() + + try: + await owner.connect("SRC", "DST") + + snapshot = await observer.snapshot() + assert _edge_exists(snapshot, "SRC", "DST") + + await owner._close_session() + await asyncio.sleep(0.05) + + snapshot = await observer.snapshot() + assert not _edge_exists(snapshot, "SRC", "DST") + finally: + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_shared_edge_survives_until_last_session_drops(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner_a = GraphContext(address, auto_start=False) + owner_b = GraphContext(address, auto_start=False) + + await owner_a.__aenter__() + await owner_b.__aenter__() + + try: + await owner_a.connect("SRC", "DST") + await owner_b.connect("SRC", "DST") + + await owner_a._close_session() + await asyncio.sleep(0.05) + + snapshot = await owner_b.snapshot() + assert _edge_exists(snapshot, "SRC", "DST") + + await owner_b._close_session() + await asyncio.sleep(0.05) + + dag = await GraphService(address).dag() + assert "DST" not in dag.graph.get("SRC", set()) + finally: + await owner_a.__aexit__(None, None, None) + await owner_b.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_session_metadata_drops_with_session(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + + await owner.__aenter__() + await observer.__aenter__() + + try: + metadata = GraphMetadata( + schema_version=1, + root_name="TEST", + components={}, + ) + await owner.register_metadata(metadata) + + owner_session_id = str(owner._session_id) + snapshot = await observer.snapshot() + assert owner_session_id in snapshot.sessions + assert snapshot.sessions[owner_session_id].metadata == metadata + + await owner._close_session() + await asyncio.sleep(0.05) + + snapshot = await observer.snapshot() + assert owner_session_id not in snapshot.sessions + finally: + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() diff --git a/tests/test_perf_ab.py b/tests/test_perf_ab.py new file mode 100644 index 00000000..3fbb935c --- /dev/null +++ b/tests/test_perf_ab.py @@ -0,0 +1,62 @@ +from ezmsg.util.perf.ab import ( + build_hotpath_command, + build_pair_order, + summarize_ab_results, +) + + +def test_build_pair_order_is_balanced_and_reproducible(): + first = build_pair_order(6, seed=123) + second = build_pair_order(6, seed=123) + + assert first == second + assert len(first) == 6 + assert first.count(("A", "B")) == 3 + assert first.count(("B", "A")) == 3 + + +def test_build_hotpath_command_contains_expected_args(tmp_path): + cmd = build_hotpath_command( + tmp_path / "out.json", + count=100, + warmup=10, + payload_sizes=[64, 256], + transports=["local", "shm"], + apis=["async", "sync"], + num_buffers=2, + quiet=True, + ) + + assert cmd[:5] == ["uv", "run", "python", "-m", "ezmsg.util.perf.hotpath"] + assert "--count" in cmd + assert "--payload-sizes" in cmd + assert "--quiet" in cmd + + +def test_summarize_ab_results_uses_b_vs_a_delta(): + paired_runs = [ + ( + {"async/shm/payload=64/buffers=1": 10.0}, + {"async/shm/payload=64/buffers=1": 12.0}, + ), + ( + {"async/shm/payload=64/buffers=1": 8.0}, + {"async/shm/payload=64/buffers=1": 9.0}, + ), + ] + + summary = summarize_ab_results( + ref_a="dev", + ref_b="CURRENT", + rounds=2, + seed=0, + paired_runs=paired_runs, + ) + + assert len(summary.cases) == 1 + case = summary.cases[0] + assert case.case_id == "async/shm/payload=64/buffers=1" + assert case.a_us_per_message_median == 9.0 + assert case.b_us_per_message_median == 10.5 + assert case.delta_pct_median > 0 + assert case.b_faster_pairs == 0 diff --git a/tests/test_perf_configs.py b/tests/test_perf_configs.py index fc7178aa..239a8868 100644 --- a/tests/test_perf_configs.py +++ b/tests/test_perf_configs.py @@ -1,4 +1,5 @@ import contextlib +from dataclasses import replace import os import tempfile from pathlib import Path @@ -6,7 +7,16 @@ import pytest from ezmsg.core.graphserver import GraphServer -from ezmsg.util.perf.impl import Communication, CONFIGS, perform_test +from ezmsg.util.perf.impl import ( + Communication, + CONFIGS, + ConfigSettings, + LoadTestSettings, + LoadTestSink, + LoadTestSource, + fanin, + perform_test, +) PERF_MAX_DURATION = 0.5 @@ -93,6 +103,26 @@ def test_fanin_perf(perf_graph_server, comm, msg_size): _run_perf_case("fanin", comm, msg_size, perf_graph_server) +def test_fanin_config_counts_all_publishers(): + settings = LoadTestSettings( + max_duration=1.0, + num_msgs=8, + dynamic_size=64, + buffers=2, + force_tcp=False, + ) + source = LoadTestSource(settings) + sink = LoadTestSink(settings) + + clients, connections = fanin( + ConfigSettings(n_clients=2, settings=settings, source=source, sink=sink) + ) + + assert len(clients) == 2 + assert len(connections) == 3 + assert sink.SETTINGS == replace(settings, num_msgs=24) + + @pytest.mark.parametrize("msg_size", PERF_MSG_SIZES, ids=lambda s: f"msg={s}") @pytest.mark.parametrize("comm", list(Communication), ids=lambda c: f"comm={c.value}") def test_relay_perf(perf_graph_server, comm, msg_size): diff --git a/tests/test_perf_hotpath.py b/tests/test_perf_hotpath.py new file mode 100644 index 00000000..9e0fcc6f --- /dev/null +++ b/tests/test_perf_hotpath.py @@ -0,0 +1,45 @@ +import pytest + +from ezmsg.util.perf.hotpath import HotPathCase, build_cases, run_hotpath_case + +from ezmsg.core.graphserver import GraphServer + + +def test_build_cases_are_sorted_by_case_id(): + cases = build_cases( + apis=["sync", "async"], + transports=["local", "tcp", "shm"], + payload_sizes=[1024, 64], + num_buffers=1, + ) + assert [case.case_id for case in cases] == sorted(case.case_id for case in cases) + assert "async/shm/payload=64/buffers=1" in {case.case_id for case in cases} + assert "async/local/payload=64/buffers=1" in {case.case_id for case in cases} + + +def test_run_hotpath_case_smoke(): + server = GraphServer() + try: + server.start(("127.0.0.1", 0)) + except PermissionError: + pytest.skip("Local socket binding is unavailable in this environment") + try: + result = run_hotpath_case( + HotPathCase( + api="sync", + transport="tcp", + payload_size=64, + num_buffers=1, + ), + count=8, + warmup=2, + samples=2, + graph_address=server.address, + ) + finally: + server.stop() + + assert result.case.case_id == "sync/tcp/payload=64/buffers=1" + assert len(result.samples_seconds) == 2 + assert all(sample > 0 for sample in result.samples_seconds) + assert result.summary.us_per_message_median > 0 diff --git a/tests/test_process_control.py b/tests/test_process_control.py new file mode 100644 index 00000000..81a22df1 --- /dev/null +++ b/tests/test_process_control.py @@ -0,0 +1,186 @@ +import asyncio +import pickle + +import pytest + +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import ProcessRegistration +from ezmsg.core.processclient import ProcessControlClient +from ezmsg.core.graphserver import GraphService +from ezmsg.core.netprotocol import Command, close_stream_writer, read_str, uint64_to_bytes + + +@pytest.mark.asyncio +async def test_process_registration_visible_in_snapshot(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + + try: + await process.register(["SYS/U1", "SYS/U2"]) + await process.update_ownership(added_units=["SYS/U3"], removed_units=["SYS/U1"]) + + snapshot = await observer.snapshot() + assert len(snapshot.processes) == 1 + + process_entry = next(iter(snapshot.processes.values())) + assert process_entry.process_id == process_key + assert process_entry.pid is not None + assert process_entry.host is not None + assert process_entry.units == ["SYS/U2", "SYS/U3"] + + finally: + await process.close() + await asyncio.sleep(0.05) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_snapshot_entry_drops_on_disconnect(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + + try: + await process.register(["SYS/U1"]) + snapshot = await observer.snapshot() + assert len(snapshot.processes) == 1 + + await process.close() + await asyncio.sleep(0.05) + + snapshot = await observer.snapshot() + assert len(snapshot.processes) == 0 + + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "command", + [ + Command.PROCESS_REGISTER, + Command.PROCESS_UPDATE_OWNERSHIP, + Command.PROCESS_SETTINGS_UPDATE, + ], +) +async def test_process_payload_parse_failures_return_error_ack(command: Command): + graph_server = GraphService().create_server() + address = graph_server.address + + reader, writer = await GraphService(address).open_connection() + try: + writer.write(Command.PROCESS.value) + await writer.drain() + _client_id = await read_str(reader) + response = await reader.read(1) + assert response == Command.COMPLETE.value + + # Non-pickled bytes intentionally trigger parse failure in process handlers. + bad_payload = b"not-a-pickle-payload" + writer.write(command.value) + writer.write(uint64_to_bytes(len(bad_payload))) + writer.write(bad_payload) + await writer.drain() + + response = await reader.read(1) + assert response == Command.ERROR.value + finally: + await close_stream_writer(writer) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_register_succeeds_after_error_ack(): + graph_server = GraphService().create_server() + address = graph_server.address + + reader, writer = await GraphService(address).open_connection() + try: + writer.write(Command.PROCESS.value) + await writer.drain() + _client_id = await read_str(reader) + response = await reader.read(1) + assert response == Command.COMPLETE.value + + bad_payload = b"not-a-pickle-payload" + writer.write(Command.PROCESS_REGISTER.value) + writer.write(uint64_to_bytes(len(bad_payload))) + writer.write(bad_payload) + await writer.drain() + response = await reader.read(1) + assert response == Command.ERROR.value + + good_payload = pickle.dumps( + ProcessRegistration( + pid=123, + host="test-host", + units=["SYS/U1"], + ) + ) + writer.write(Command.PROCESS_REGISTER.value) + writer.write(uint64_to_bytes(len(good_payload))) + writer.write(good_payload) + await writer.drain() + response = await reader.read(1) + assert response == Command.COMPLETE.value + finally: + await close_stream_writer(writer) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_register_rejects_unit_ownership_collision(): + graph_server = GraphService().create_server() + address = graph_server.address + + process_a = ProcessControlClient(address) + process_b = ProcessControlClient(address) + await process_a.connect() + await process_b.connect() + + try: + await process_a.register(["SYS/U1"]) + with pytest.raises(RuntimeError, match="PROCESS_REGISTER"): + await process_b.register(["SYS/U1"]) + finally: + await process_a.close() + await process_b.close() + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_update_ownership_rejects_unit_ownership_collision(): + graph_server = GraphService().create_server() + address = graph_server.address + + process_a = ProcessControlClient(address) + process_b = ProcessControlClient(address) + await process_a.connect() + await process_b.connect() + + try: + await process_a.register(["SYS/U1"]) + await process_b.register(["SYS/U2"]) + with pytest.raises(RuntimeError, match="PROCESS_UPDATE_OWNERSHIP"): + await process_b.update_ownership(added_units=["SYS/U1"]) + finally: + await process_a.close() + await process_b.close() + graph_server.stop() diff --git a/tests/test_process_routing.py b/tests/test_process_routing.py new file mode 100644 index 00000000..2402466d --- /dev/null +++ b/tests/test_process_routing.py @@ -0,0 +1,150 @@ +import asyncio + +import pytest + +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import ( + ProcessControlErrorCode, + ProcessControlRequest, + ProcessControlResponse, +) +from ezmsg.core.graphserver import GraphService +from ezmsg.core.processclient import ProcessControlClient + + +@pytest.mark.asyncio +async def test_process_routing_roundtrip(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U1"]) + + async def handler(request: ProcessControlRequest) -> ProcessControlResponse: + assert request.unit_address == "SYS/U1" + assert request.operation == "ECHO" + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + payload=request.payload, + ) + + process.set_request_handler(handler) + + try: + response = await observer.process_request( + "SYS/U1", + "ECHO", + payload=b"hello", + timeout=1.0, + ) + assert response.ok + assert response.payload == b"hello" + assert response.process_id == process_key + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_routing_builtin_ping_and_stats(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U1", "SYS/U2"]) + await process.update_ownership(removed_units=["SYS/U2"], added_units=["SYS/U3"]) + + try: + ping = await observer.process_ping("SYS/U1", timeout=1.0) + assert ping.process_id == process_key + assert ping.pid > 0 + assert ping.host + assert ping.timestamp > 0.0 + + stats = await observer.process_stats("SYS/U1", timeout=1.0) + assert stats.process_id == process_key + assert stats.pid > 0 + assert stats.host + assert stats.owned_units == ["SYS/U1", "SYS/U3"] + assert stats.timestamp > 0.0 + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_routing_missing_owner_returns_error(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + try: + response = await observer.process_request( + "SYS/UNKNOWN", + "PING", + payload=b"", + timeout=0.25, + ) + assert not response.ok + assert response.error is not None + assert "No process owns unit" in response.error + assert response.error_code == ProcessControlErrorCode.UNROUTABLE_UNIT + finally: + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_routing_timeout_returns_error(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U2"]) + + block = asyncio.Event() + + async def blocking_handler(_request: ProcessControlRequest) -> ProcessControlResponse: + await block.wait() + return ProcessControlResponse(request_id="", ok=False) + + process.set_request_handler(blocking_handler) + + try: + response = await observer.process_request( + "SYS/U2", + "SLOW", + timeout=0.05, + ) + assert not response.ok + assert response.error is not None + assert "Timed out waiting for process response" in response.error + assert response.error_code == ProcessControlErrorCode.TIMEOUT + assert response.process_id == process_key + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() diff --git a/tests/test_profiling_api.py b/tests/test_profiling_api.py new file mode 100644 index 00000000..5c7d2859 --- /dev/null +++ b/tests/test_profiling_api.py @@ -0,0 +1,636 @@ +import asyncio + +import pytest + +from uuid import uuid4 + +from ezmsg.core import profiling as profiling_core +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import ( + ProcessControlErrorCode, + ProfileChannelType, + ProfilingStreamControl, + ProfilingTraceControl, +) +from ezmsg.core.graphserver import GraphService +from ezmsg.core.processclient import ProcessControlClient + + +def test_profiling_snapshot_uses_counter_deltas_between_snapshots( + monkeypatch: pytest.MonkeyPatch, +): + registry = profiling_core.ProfileRegistry() + publisher = registry.register_publisher(uuid4(), "TOPIC_IDLE", 4) + subscriber = registry.register_subscriber(uuid4(), "TOPIC_IDLE") + + now_ns = {"value": int(0.0)} + monkeypatch.setattr(profiling_core, "PROFILE_TIME", lambda: now_ns["value"]) + + publisher.record_publish(inflight=0) + publisher.record_publish(inflight=1) + subscriber.record_receive(ProfileChannelType.LOCAL) + + first = registry.snapshot() + pub_first = next(iter(first.publishers.values())) + sub_first = next(iter(first.subscribers.values())) + assert first.window_seconds == 0.0 + assert pub_first.messages_published_total == 2 + assert pub_first.messages_published_window == 0 + assert pub_first.publish_rate_hz_window == 0.0 + assert pub_first.inflight_messages_current == 1 + assert sub_first.messages_received_total == 1 + assert sub_first.messages_received_window == 0 + assert sub_first.channel_kind_last == ProfileChannelType.LOCAL + + now_ns["value"] = int(1.0e9) + publisher.record_publish(inflight=0) + subscriber.record_receive(ProfileChannelType.SHM) + + second = registry.snapshot() + pub_second = next(iter(second.publishers.values())) + sub_second = next(iter(second.subscribers.values())) + assert second.window_seconds == pytest.approx(1.0) + assert pub_second.messages_published_total == 3 + assert pub_second.messages_published_window == 1 + assert pub_second.publish_rate_hz_window == pytest.approx(1.0) + assert pub_second.inflight_messages_current == 0 + assert sub_second.messages_received_total == 2 + assert sub_second.messages_received_window == 1 + assert sub_second.channel_kind_last == ProfileChannelType.SHM + + +def test_subscriber_trace_sampling_uses_one_decision_per_message(): + registry = profiling_core.ProfileRegistry() + subscriber = registry.register_subscriber(uuid4(), "TOPIC_SAMPLE") + registry.set_trace_control( + ProfilingTraceControl( + enabled=True, + sample_mod=3, + subscriber_topics=["TOPIC_SAMPLE"], + metrics=["lease_time_ns", "user_span_ns"], + ) + ) + + for msg_seq in range(7): + sampled = subscriber.begin_message(ProfileChannelType.LOCAL) + subscriber.record_lease_time( + ProfileChannelType.LOCAL, + 100 + msg_seq, + msg_seq=msg_seq, + sampled=sampled, + ) + subscriber.record_user_span( + 200 + msg_seq, + "taskA", + msg_seq=msg_seq, + sampled=sampled, + ) + + batch = registry.trace_batch(max_samples=100) + lease_samples = [sample for sample in batch.samples if sample.metric == "lease_time_ns"] + user_span_samples = [sample for sample in batch.samples if sample.metric == "user_span_ns"] + + assert [sample.sample_seq for sample in lease_samples] == [2, 5] + assert [sample.sample_seq for sample in user_span_samples] == [2, 5] + + +@pytest.mark.asyncio +async def test_process_profiling_snapshot_collects_pub_sub_metrics(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U1"]) + + pub = await ctx.publisher("TOPIC_PROF") + sub = await ctx.subscriber("TOPIC_PROF") + + try: + for idx in range(8): + await pub.broadcast(idx) + async with sub.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + snap = await ctx.process_profiling_snapshot("SYS/U1", timeout=1.0) + assert snap.process_id == process_key + assert snap.window_seconds >= 0.0 + assert len(snap.publishers) >= 1 + assert len(snap.subscribers) >= 1 + + pub_metrics = next( + pub for pub in snap.publishers.values() if pub.topic == "TOPIC_PROF" + ) + assert pub_metrics.messages_published_total >= 8 + assert pub_metrics.num_buffers > 0 + assert pub_metrics.inflight_messages_current >= 0 + + sub_metrics = next( + sub for sub in snap.subscribers.values() if sub.topic == "TOPIC_PROF" + ) + assert sub_metrics.messages_received_total >= 8 + assert sub_metrics.channel_kind_last != ProfileChannelType.UNKNOWN + finally: + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_connect_does_not_clear_preexisting_profile_metrics(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + pub = await ctx.publisher("TOPIC_PRECONNECT") + sub = await ctx.subscriber("TOPIC_PRECONNECT") + for idx in range(6): + await pub.broadcast(idx) + async with sub.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/U_PRE"]) + + try: + snap = await ctx.process_profiling_snapshot("SYS/U_PRE", timeout=1.0) + assert len(snap.publishers) >= 1 + assert len(snap.subscribers) >= 1 + pub_metrics = next( + pub for pub in snap.publishers.values() if pub.topic == "TOPIC_PRECONNECT" + ) + sub_metrics = next( + sub for sub in snap.subscribers.values() if sub.topic == "TOPIC_PRECONNECT" + ) + assert pub_metrics.messages_published_total >= 6 + assert sub_metrics.messages_received_total >= 6 + finally: + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_profiling_trace_control_and_batch(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U2"]) + # Keep direct batch assertions deterministic by preventing the background + # trace-push loop from draining the local trace buffer during this test. + process._trace_push_interval_s = 60.0 + + pub = await ctx.publisher("TOPIC_TRACE") + sub = await ctx.subscriber("TOPIC_TRACE") + + try: + response = await ctx.process_set_profiling_trace( + "SYS/U2", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_topics=["TOPIC_TRACE"], + subscriber_topics=["TOPIC_TRACE"], + ), + timeout=1.0, + ) + assert response.ok + + for idx in range(5): + await pub.broadcast(idx) + async with sub.recv_zero_copy() as _msg: + span_start_ns = sub.begin_profile() + try: + await asyncio.sleep(0) + finally: + sub.end_profile(span_start_ns, "taskA") + + batch = await ctx.process_profiling_trace_batch( + "SYS/U2", max_samples=200, timeout=1.0 + ) + assert batch.process_id == process_key + assert len(batch.samples) > 0 + assert "publish_delta_ns" in {sample.metric for sample in batch.samples} + assert "lease_time_ns" in {sample.metric for sample in batch.samples} + assert "user_span_ns" in {sample.metric for sample in batch.samples} + + disable_response = await ctx.process_set_profiling_trace( + "SYS/U2", + ProfilingTraceControl(enabled=False), + timeout=1.0, + ) + assert disable_response.ok + finally: + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_profiling_snapshot_all_and_unroutable_error_code(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U3"]) + + try: + snapshots = await ctx.profiling_snapshot_all(timeout_per_process=0.5) + assert process_key in snapshots + assert snapshots[process_key].process_id == process_key + + response = await ctx.process_request( + "SYS/MISSING", + "GET_PROFILING_SNAPSHOT", + timeout=0.2, + ) + assert not response.ok + assert response.error_code == ProcessControlErrorCode.UNROUTABLE_UNIT + finally: + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_profiling_trace_subscription_push(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U4"]) + + pub = await ctx.publisher("TOPIC_STREAM") + sub = await ctx.subscriber("TOPIC_STREAM") + stream = None + + try: + response = await ctx.process_set_profiling_trace( + "SYS/U4", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_topics=["TOPIC_STREAM"], + subscriber_topics=["TOPIC_STREAM"], + ), + timeout=1.0, + ) + assert response.ok + + stream = ctx.subscribe_profiling_trace( + ProfilingStreamControl(interval=0.02, max_samples=256) + ) + + for idx in range(8): + await pub.broadcast(idx) + async with sub.recv_zero_copy() as _msg: + span_start_ns = sub.begin_profile() + try: + await asyncio.sleep(0) + finally: + sub.end_profile(span_start_ns, "taskA") + + batch = await asyncio.wait_for(anext(stream), timeout=1.0) + assert batch.timestamp > 0.0 + assert process_key in batch.batches + process_batch = batch.batches[process_key] + assert process_batch.process_id == process_key + assert len(process_batch.samples) > 0 + finally: + if stream is not None: + await stream.aclose() + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_profiling_trace_control_endpoint_metric_and_ttl(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/U5"]) + # This test reads batches directly from the process, so disable the + # automatic push loop long enough that it cannot race and drain samples. + process._trace_push_interval_s = 60.0 + + pub_a = await ctx.publisher("TOPIC_A") + sub_a = await ctx.subscriber("TOPIC_A") + pub_b = await ctx.publisher("TOPIC_B") + sub_b = await ctx.subscriber("TOPIC_B") + + try: + # Warm up and discover endpoint IDs for precise filter targeting. + for idx in range(3): + await pub_a.broadcast(idx) + async with sub_a.recv_zero_copy() as _msg: + await asyncio.sleep(0) + await pub_b.broadcast(idx) + async with sub_b.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + snapshot = await ctx.process_profiling_snapshot("SYS/U5", timeout=1.0) + pub_a_endpoint = next( + pub.endpoint_id + for pub in snapshot.publishers.values() + if pub.topic == "TOPIC_A" + ) + + response = await ctx.process_set_profiling_trace( + "SYS/U5", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_endpoint_ids=[pub_a_endpoint], + metrics=["publish_delta_ns"], + ), + timeout=1.0, + ) + assert response.ok + + for idx in range(8): + await pub_a.broadcast(idx) + async with sub_a.recv_zero_copy() as _msg: + await asyncio.sleep(0) + await pub_b.broadcast(idx) + async with sub_b.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + batch = await ctx.process_profiling_trace_batch( + "SYS/U5", max_samples=512, timeout=1.0 + ) + assert len(batch.samples) > 0 + assert all(sample.metric == "publish_delta_ns" for sample in batch.samples) + assert all(sample.endpoint_id == pub_a_endpoint for sample in batch.samples) + + ttl_response = await ctx.process_set_profiling_trace( + "SYS/U5", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_endpoint_ids=[pub_a_endpoint], + metrics=["publish_delta_ns"], + ttl_seconds=0.01, + ), + timeout=1.0, + ) + assert ttl_response.ok + await asyncio.sleep(0.03) + + for idx in range(3): + await pub_a.broadcast(idx) + async with sub_a.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + expired_batch = await ctx.process_profiling_trace_batch( + "SYS/U5", max_samples=512, timeout=1.0 + ) + assert len(expired_batch.samples) == 0 + finally: + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_profiling_trace_subscription_stream_control(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process_a = ProcessControlClient(address) + await process_a.connect() + assert process_a.client_id is not None + process_a_key = process_a.client_id + await process_a.register(["SYS/U6"]) + + stream = None + try: + stream = ctx.subscribe_profiling_trace( + ProfilingStreamControl( + interval=0.02, + max_samples=64, + process_ids=[process_a_key], + include_empty_batches=True, + ) + ) + batch = await asyncio.wait_for(anext(stream), timeout=1.0) + assert process_a_key in batch.batches + assert len(batch.batches) == 1 + finally: + if stream is not None: + await stream.aclose() + await process_a.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_profiling_trace_subscription_does_not_starve_peer_subscribers(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U7"]) + + pub = await ctx.publisher("TOPIC_STREAM_MULTI") + sub = await ctx.subscriber("TOPIC_STREAM_MULTI") + stream_a = None + stream_b = None + + try: + response = await ctx.process_set_profiling_trace( + "SYS/U7", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_topics=["TOPIC_STREAM_MULTI"], + subscriber_topics=["TOPIC_STREAM_MULTI"], + ), + timeout=1.0, + ) + assert response.ok + + control = ProfilingStreamControl(interval=0.02, max_samples=256) + stream_a = ctx.subscribe_profiling_trace(control) + stream_b = ctx.subscribe_profiling_trace(control) + + for idx in range(12): + await pub.broadcast(idx) + async with sub.recv_zero_copy() as _msg: + span_start_ns = sub.begin_profile() + try: + await asyncio.sleep(0) + finally: + sub.end_profile(span_start_ns, "taskA") + + batch_a = await asyncio.wait_for(anext(stream_a), timeout=1.0) + batch_b = await asyncio.wait_for(anext(stream_b), timeout=1.0) + + assert process_key in batch_a.batches + assert process_key in batch_b.batches + assert len(batch_a.batches[process_key].samples) > 0 + assert len(batch_b.batches[process_key].samples) > 0 + finally: + if stream_a is not None: + await stream_a.aclose() + if stream_b is not None: + await stream_b.aclose() + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_profiling_trace_batch_interleaves_publisher_and_subscriber_samples(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/U8"]) + # Avoid races with the automatic graph-server push path when reading the + # process-local trace batch directly in this test. + process._trace_push_interval_s = 60.0 + + pub = await ctx.publisher("TOPIC_TRACE_MIX") + sub = await ctx.subscriber("TOPIC_TRACE_MIX") + + try: + response = await ctx.process_set_profiling_trace( + "SYS/U8", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_topics=["TOPIC_TRACE_MIX"], + subscriber_topics=["TOPIC_TRACE_MIX"], + metrics=["publish_delta_ns", "lease_time_ns"], + ), + timeout=1.0, + ) + assert response.ok + + for idx in range(64): + await pub.broadcast(idx) + async with sub.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + batch = await ctx.process_profiling_trace_batch( + "SYS/U8", max_samples=32, timeout=1.0 + ) + metrics = {sample.metric for sample in batch.samples} + assert "publish_delta_ns" in metrics + assert "lease_time_ns" in metrics + finally: + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_profiling_trace_control_change_clears_stale_trace_samples(): + graph_server = GraphService().create_server() + address = graph_server.address + + ctx = GraphContext(address, auto_start=False) + await ctx.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/U9"]) + # This test validates local batch contents after a control change, so keep + # the background push loop from draining new samples before the assertion. + process._trace_push_interval_s = 60.0 + + pub_old = await ctx.publisher("TOPIC_TRACE_OLD") + sub_old = await ctx.subscriber("TOPIC_TRACE_OLD") + pub_new = await ctx.publisher("TOPIC_TRACE_NEW") + sub_new = await ctx.subscriber("TOPIC_TRACE_NEW") + + try: + old_response = await ctx.process_set_profiling_trace( + "SYS/U9", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_topics=["TOPIC_TRACE_OLD"], + metrics=["publish_delta_ns"], + ), + timeout=1.0, + ) + assert old_response.ok + + for idx in range(12): + await pub_old.broadcast(idx) + async with sub_old.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + new_response = await ctx.process_set_profiling_trace( + "SYS/U9", + ProfilingTraceControl( + enabled=True, + sample_mod=1, + publisher_topics=["TOPIC_TRACE_NEW"], + metrics=["publish_delta_ns"], + ), + timeout=1.0, + ) + assert new_response.ok + + for idx in range(8): + await pub_new.broadcast(idx) + async with sub_new.recv_zero_copy() as _msg: + await asyncio.sleep(0) + + batch = await ctx.process_profiling_trace_batch( + "SYS/U9", max_samples=256, timeout=1.0 + ) + assert len(batch.samples) > 0 + assert all(sample.topic == "TOPIC_TRACE_NEW" for sample in batch.samples) + finally: + await process.close() + await ctx.__aexit__(None, None, None) + graph_server.stop() diff --git a/tests/test_pubclient.py b/tests/test_pubclient.py new file mode 100644 index 00000000..27048b4d --- /dev/null +++ b/tests/test_pubclient.py @@ -0,0 +1,203 @@ +from contextlib import contextmanager +from uuid import uuid4 + +import pytest + +from ezmsg.core.netprotocol import Address, Command +from ezmsg.core.pubclient import ( + ALLOW_LOCAL_ENV, + FORCE_TCP_ENV, + PubChannelInfo, + Publisher, +) + + +class DummyLocalChannel: + def __init__(self) -> None: + self.calls: list[tuple[int, object]] = [] + + def put_local(self, msg_id: int, obj: object) -> None: + self.calls.append((msg_id, obj)) + + +class DummyWriter: + def __init__(self) -> None: + self.buffer: list[bytes] = [] + + def write(self, data: bytes) -> None: + self.buffer.append(data) + + async def drain(self) -> None: + return None + + +class DummyShm: + def __init__(self, num_buffers: int, buf_size: int = 65536) -> None: + self.name = "dummy-shm" + self.buf_size = buf_size + self._buffers = [bytearray(buf_size) for _ in range(num_buffers)] + + @contextmanager + def buffer(self, idx: int, readonly: bool = False): + del readonly + yield memoryview(self._buffers[idx]) + + +def _make_publisher( + *, + force_tcp: bool | None, + allow_local: bool | None, + channel_pid: int, + shm_ok: bool = True, +) -> tuple[Publisher, DummyLocalChannel, DummyWriter]: + pub = Publisher( + id=uuid4(), + topic="/TEST", + shm=DummyShm(num_buffers=2), + graph_address=Address("127.0.0.1", 25978), + num_buffers=2, + force_tcp=force_tcp, + allow_local=allow_local, + _guard=Publisher._SENTINEL, + ) + pub._running.set() + + local_channel = DummyLocalChannel() + writer = DummyWriter() + channel = PubChannelInfo( + id=uuid4(), + writer=writer, + pub_id=pub.id, + pid=channel_pid, + shm_ok=shm_ok, + ) + pub._channels[channel.id] = channel + pub._local_channel = local_channel # type: ignore[assignment] + return pub, local_channel, writer + + +@pytest.mark.asyncio +async def test_broadcast_same_process_prefers_local_fast_path(): + pub, local_channel, writer = _make_publisher( + force_tcp=False, + allow_local=True, + channel_pid=0, + ) + pub.pid = 0 + + await pub.broadcast(b"payload") + + assert local_channel.calls == [(0, b"payload")] + assert writer.buffer == [] + + +@pytest.mark.asyncio +async def test_broadcast_same_process_can_force_shm_path(): + pub, local_channel, writer = _make_publisher( + force_tcp=False, + allow_local=False, + channel_pid=0, + ) + pub.pid = 0 + + await pub.broadcast(b"payload") + + assert local_channel.calls == [] + assert writer.buffer + assert writer.buffer[0].startswith(Command.TX_SHM.value) + + +@pytest.mark.asyncio +async def test_broadcast_same_process_can_force_tcp_path(): + pub, local_channel, writer = _make_publisher( + force_tcp=True, + allow_local=False, + channel_pid=0, + ) + pub.pid = 0 + + await pub.broadcast(b"payload") + + assert local_channel.calls == [] + assert writer.buffer + assert writer.buffer[0].startswith(Command.TX_TCP.value) + + +def test_force_tcp_disables_allow_local_from_env(monkeypatch, caplog): + monkeypatch.setenv(ALLOW_LOCAL_ENV, "1") + with caplog.at_level("INFO"): + pub, _, _ = _make_publisher( + force_tcp=True, + allow_local=None, + channel_pid=0, + ) + + assert pub._allow_local is False + assert "force_tcp=True disables local delivery" in caplog.text + + +def test_force_tcp_disables_explicit_allow_local(caplog): + with caplog.at_level("INFO"): + pub, _, _ = _make_publisher( + force_tcp=True, + allow_local=True, + channel_pid=0, + ) + + assert pub._allow_local is False + assert "force_tcp=True disables local delivery" in caplog.text + + +def test_force_tcp_uses_env_default_when_none(monkeypatch): + monkeypatch.setenv(FORCE_TCP_ENV, "1") + pub, _, _ = _make_publisher( + force_tcp=None, + allow_local=False, + channel_pid=0, + ) + + assert pub._force_tcp is True + + +def test_explicit_force_tcp_false_overrides_env(monkeypatch): + monkeypatch.setenv(FORCE_TCP_ENV, "1") + pub, _, _ = _make_publisher( + force_tcp=False, + allow_local=False, + channel_pid=0, + ) + + assert pub._force_tcp is False + + +@pytest.mark.asyncio +async def test_broadcast_same_process_uses_env_default_when_allow_local_is_none(monkeypatch): + monkeypatch.setenv(ALLOW_LOCAL_ENV, "0") + pub, local_channel, writer = _make_publisher( + force_tcp=False, + allow_local=None, + channel_pid=0, + ) + pub.pid = 0 + + await pub.broadcast(b"payload") + + assert local_channel.calls == [] + assert writer.buffer + assert writer.buffer[0].startswith(Command.TX_SHM.value) + + +@pytest.mark.asyncio +async def test_broadcast_same_process_explicit_allow_local_overrides_env(monkeypatch): + monkeypatch.setenv(ALLOW_LOCAL_ENV, "0") + pub, local_channel, writer = _make_publisher( + force_tcp=False, + allow_local=True, + channel_pid=0, + ) + pub.pid = 0 + + await pub.broadcast(b"payload") + + assert local_channel.calls == [(0, b"payload")] + assert writer.buffer == [] diff --git a/tests/test_settings_api.py b/tests/test_settings_api.py new file mode 100644 index 00000000..2eb273dc --- /dev/null +++ b/tests/test_settings_api.py @@ -0,0 +1,538 @@ +import asyncio +import pickle +import time +from dataclasses import dataclass + +import pytest + +import ezmsg.core as ez +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import ( + ComponentMetadata, + DynamicSettingsMetadata, + GraphMetadata, + ProcessControlErrorCode, + ProcessControlResponse, + SettingsFieldUpdateRequest, + SettingsEventType, + SettingsSnapshotValue, +) +from ezmsg.core.graphserver import GraphService +from ezmsg.core.processclient import ProcessControlClient + + +def _metadata_with_component(component_address: str) -> GraphMetadata: + return GraphMetadata( + schema_version=1, + root_name="SYS", + components={ + component_address: ComponentMetadata( + address=component_address, + name="UNIT", + component_type="example.Unit", + settings_type="example.Settings", + initial_settings=(None, {"alpha": 1}), + dynamic_settings=DynamicSettingsMetadata( + enabled=True, + input_topic=f"{component_address}/INPUT_SETTINGS", + settings_type="example.Settings", + ), + settings_schema=None, + ) + }, + ) + + +@pytest.mark.asyncio +async def test_settings_snapshot_and_events_from_metadata_registration(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + + await owner.__aenter__() + await observer.__aenter__() + + try: + component_address = "SYS/UNIT_A" + await owner.register_metadata(_metadata_with_component(component_address)) + + settings = await observer.settings_snapshot() + assert component_address in settings + assert settings[component_address].repr_value == {"alpha": 1} + assert settings[component_address].structured_value == {"alpha": 1} + assert settings[component_address].settings_schema is None + + events = await observer.settings_events(after_seq=0) + matching = [ + event + for event in events + if event.component_address == component_address + and event.event_type == SettingsEventType.INITIAL_SETTINGS + ] + assert matching + assert matching[-1].source_session_id == str(owner._session_id) + finally: + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@dataclass +class _SettingsMsg: + gain: int + + +class _SettingsSource(ez.Unit): + OUTPUT = ez.OutputStream(_SettingsMsg) + + @ez.publisher(OUTPUT) + async def emit(self): + yield self.OUTPUT, _SettingsMsg(gain=7) + raise ez.Complete + + +class _SettingsSink(ez.Unit): + INPUT_SETTINGS = ez.InputStream(_SettingsMsg) + + @ez.subscriber(INPUT_SETTINGS) + async def on_settings(self, msg: _SettingsMsg) -> None: + raise ez.NormalTermination + + +class _SettingsSystem(ez.Collection): + SRC = _SettingsSource() + SINK = _SettingsSink() + + def network(self) -> ez.NetworkDefinition: + return ((self.SRC.OUTPUT, self.SINK.INPUT_SETTINGS),) + + +class _SettingsOnlySystem(ez.Collection): + SINK = _SettingsSink() + + def network(self) -> ez.NetworkDefinition: + return () + + +def test_input_settings_hook_reports_to_graphserver(): + graph_server = GraphService().create_server() + address = graph_server.address + try: + ez.run(components={"SYS": _SettingsSystem()}, graph_address=address, force_single_process=True) + + async def observe() -> None: + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + try: + settings = await observer.settings_snapshot() + sink_address = "SYS/SINK" + # Process-owned settings are cleaned up when the process exits. + assert sink_address not in settings + + events = await observer.settings_events(after_seq=0) + matching = [ + event + for event in events + if event.component_address == sink_address + and event.event_type == SettingsEventType.SETTINGS_UPDATED + and event.value.repr_value == {"gain": 7} + ] + assert matching + latest = matching[-1].value + assert latest.structured_value == {"gain": 7} + assert latest.settings_schema is not None + schema = latest.settings_schema + assert schema.provider == "dataclass" + assert any( + field.name == "gain" and "int" in field.field_type.lower() + for field in schema.fields + ) + finally: + await observer.__aexit__(None, None, None) + + asyncio.run(observe()) + finally: + graph_server.stop() + + +@pytest.mark.asyncio +async def test_graphcontext_update_settings_via_input_settings_topic(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + run_task = asyncio.create_task( + asyncio.to_thread( + ez.run, + components={"SYS": _SettingsOnlySystem()}, + graph_address=address, + force_single_process=True, + ) + ) + + try: + for _ in range(40): + try: + await observer.settings_input_topic("SYS/SINK") + break + except RuntimeError: + await asyncio.sleep(0.05) + else: + raise AssertionError("Timed out waiting for dynamic settings metadata") + + await observer.update_settings("SYS/SINK", _SettingsMsg(gain=11)) + await asyncio.wait_for(run_task, timeout=5.0) + + settings = await observer.settings_snapshot() + assert "SYS/SINK" not in settings + + events = await observer.settings_events(after_seq=0) + matching = [ + event + for event in events + if event.component_address == "SYS/SINK" + and event.event_type == SettingsEventType.SETTINGS_UPDATED + and event.value.repr_value == {"gain": 11} + ] + assert matching + + finally: + if not run_task.done(): + await asyncio.wait_for(run_task, timeout=5.0) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_graphcontext_update_setting_field_routes_to_process(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/SINK"]) + + try: + async def handler(request) -> ProcessControlResponse: + assert request.operation == "UPDATE_SETTING_FIELD" + assert request.payload is not None + update = pickle.loads(request.payload) + assert isinstance(update, SettingsFieldUpdateRequest) + assert update.field_path == "gain" + assert update.value == 11 + return ProcessControlResponse( + request_id=request.request_id, + ok=True, + payload=pickle.dumps( + SettingsSnapshotValue(serialized=None, repr_value={"gain": 11}) + ), + ) + + process.set_request_handler(handler) + + patched = await observer.update_setting("SYS/SINK", "gain", 11) + assert patched.repr_value == {"gain": 11} + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_graphcontext_update_setting_waits_and_propagates_process_failure(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/SINK"]) + + try: + async def handler(request) -> ProcessControlResponse: + assert request.operation == "UPDATE_SETTING_FIELD" + await asyncio.sleep(0.05) + return ProcessControlResponse( + request_id=request.request_id, + ok=False, + error="Simulated publish failure", + error_code=ProcessControlErrorCode.HANDLER_ERROR, + ) + + process.set_request_handler(handler) + + start = time.perf_counter() + with pytest.raises(RuntimeError, match="Simulated publish failure"): + await observer.update_setting("SYS/SINK", "gain", 99, timeout=1.0) + elapsed = time.perf_counter() - start + assert elapsed >= 0.04 + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_reported_settings_update_visible_in_snapshot_and_events(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + assert process.client_id is not None + process_key = process.client_id + + try: + await process.register(["SYS/UNIT_B"]) + await process.report_settings_update( + component_address="SYS/UNIT_B", + value=SettingsSnapshotValue(serialized=None, repr_value={"gain": 2}), + ) + + settings = await observer.settings_snapshot() + assert settings["SYS/UNIT_B"].repr_value == {"gain": 2} + + events = await observer.settings_events(after_seq=0) + matching = [ + event + for event in events + if event.component_address == "SYS/UNIT_B" + and event.event_type == SettingsEventType.SETTINGS_UPDATED + ] + assert matching + assert matching[-1].source_process_id == process_key + + stream = observer.subscribe_settings_events(after_seq=0) + streamed = await asyncio.wait_for(anext(stream), timeout=1.0) + assert streamed.component_address == "SYS/UNIT_B" + await stream.aclose() + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_session_owned_settings_removed_when_session_drops(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + + await owner.__aenter__() + await observer.__aenter__() + + try: + component_address = "SYS/UNIT_C" + await owner.register_metadata(_metadata_with_component(component_address)) + settings = await observer.settings_snapshot() + assert component_address in settings + + await owner._close_session() + await asyncio.sleep(0.05) + + settings = await observer.settings_snapshot() + assert component_address not in settings + finally: + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_metadata_registration_rejects_component_address_collision(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner_a = GraphContext(address, auto_start=False) + owner_b = GraphContext(address, auto_start=False) + + await owner_a.__aenter__() + await owner_b.__aenter__() + + try: + component_address = "SYS/UNIT_COLLIDE" + metadata = _metadata_with_component(component_address) + await owner_a.register_metadata(metadata) + with pytest.raises( + RuntimeError, + match="component address collision\\(s\\): SYS/UNIT_COLLIDE", + ): + await owner_b.register_metadata(metadata) + finally: + await owner_a.__aexit__(None, None, None) + await owner_b.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_owned_settings_removed_when_process_disconnects_without_session_owner(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/UNIT_ORPHAN"]) + + try: + await process.report_settings_update( + component_address="SYS/UNIT_ORPHAN", + value=SettingsSnapshotValue(serialized=None, repr_value={"gain": 5}), + ) + settings = await observer.settings_snapshot() + assert settings["SYS/UNIT_ORPHAN"].repr_value == {"gain": 5} + finally: + await process.close() + + await asyncio.sleep(0.05) + settings = await observer.settings_snapshot() + assert "SYS/UNIT_ORPHAN" not in settings + + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_disconnect_restores_metadata_initial_settings(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + await owner.__aenter__() + await observer.__aenter__() + await owner.register_metadata(_metadata_with_component("SYS/UNIT_RESTORE")) + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/UNIT_RESTORE"]) + + try: + await process.report_settings_update( + component_address="SYS/UNIT_RESTORE", + value=SettingsSnapshotValue(serialized=None, repr_value={"alpha": 9}), + ) + settings = await observer.settings_snapshot() + assert settings["SYS/UNIT_RESTORE"].repr_value == {"alpha": 9} + finally: + await process.close() + + await asyncio.sleep(0.05) + settings = await observer.settings_snapshot() + assert settings["SYS/UNIT_RESTORE"].repr_value == {"alpha": 1} + + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_session_drop_preserves_live_process_owned_settings(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + await owner.__aenter__() + await observer.__aenter__() + await owner.register_metadata(_metadata_with_component("SYS/UNIT_LIVE")) + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/UNIT_LIVE"]) + + try: + await process.report_settings_update( + component_address="SYS/UNIT_LIVE", + value=SettingsSnapshotValue(serialized=None, repr_value={"alpha": 9}), + ) + settings = await observer.settings_snapshot() + assert settings["SYS/UNIT_LIVE"].repr_value == {"alpha": 9} + + await owner._close_session() + await asyncio.sleep(0.05) + + settings = await observer.settings_snapshot() + assert settings["SYS/UNIT_LIVE"].repr_value == {"alpha": 9} + finally: + await process.close() + + await asyncio.sleep(0.05) + settings = await observer.settings_snapshot() + assert "SYS/UNIT_LIVE" not in settings + + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_settings_update_rejected_for_unowned_component(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + await process.connect() + await process.register(["SYS/UNIT_OWNED"]) + + try: + with pytest.raises(RuntimeError, match="Process control command failed"): + await process.report_settings_update( + component_address="SYS/UNIT_UNOWNED", + value=SettingsSnapshotValue(serialized=None, repr_value={"gain": 7}), + ) + settings = await observer.settings_snapshot() + assert "SYS/UNIT_UNOWNED" not in settings + finally: + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_process_settings_update_requires_completed_process_registration(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + await owner.__aenter__() + await observer.__aenter__() + await owner.register_metadata(_metadata_with_component("SYS/UNIT_PENDING")) + + process = ProcessControlClient(address) + await process.connect() + + try: + with pytest.raises(RuntimeError, match="Process control command failed"): + await process.report_settings_update( + component_address="SYS/UNIT_PENDING", + value=SettingsSnapshotValue(serialized=None, repr_value={"alpha": 7}), + ) + + settings = await observer.settings_snapshot() + assert settings["SYS/UNIT_PENDING"].repr_value == {"alpha": 1} + finally: + await process.close() + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() diff --git a/tests/test_subclient.py b/tests/test_subclient.py index d4750790..20c5ea27 100644 --- a/tests/test_subclient.py +++ b/tests/test_subclient.py @@ -5,6 +5,7 @@ import pytest from ezmsg.core.subclient import Subscriber +from ezmsg.core.graphmeta import ProfileChannelType from ezmsg.core.netprotocol import Command, encode_str from ezmsg.core import channelmanager as channelmanager_module from ezmsg.core import subclient as subclient_module @@ -19,6 +20,7 @@ def __init__(self): self.waited = False self.topic = "test" self.num_buffers = 8 + self.channel_kind = ProfileChannelType.LOCAL def register_client(self, client_id, queue, local_backpressure=None): self.clients[client_id] = queue diff --git a/tests/test_topics.py b/tests/test_topics.py new file mode 100644 index 00000000..78531bf1 --- /dev/null +++ b/tests/test_topics.py @@ -0,0 +1,189 @@ +import pytest + +import ezmsg.core as ez + +from ezmsg.core.backend import ExecutionContext +from ezmsg.core.graphmeta import ( + CollectionMetadata, + InputRelayMetadata, + InputStreamMetadata, + OutputStreamMetadata, + OutputTopicMetadata, + UnitMetadata, +) + + +@pytest.mark.parametrize( + "endpoint_factory", + [ + lambda: ez.Topic(int), + lambda: ez.InputTopic(int), + lambda: ez.OutputTopic(int), + lambda: ez.InputRelay(int), + lambda: ez.OutputRelay(int), + ], +) +def test_unit_rejects_topic_endpoints(endpoint_factory): + with pytest.raises(TypeError, match="Units may only declare InputStream"): + + class BadUnit(ez.Unit): + ENDPOINT = endpoint_factory() + + +def test_collection_stream_endpoint_warns_futurewarning(): + with pytest.warns(FutureWarning, match="deprecated"): + + class LegacyCollection(ez.Collection): + INPUT = ez.InputStream(int) + + +class _Source(ez.Unit): + OUTPUT = ez.OutputStream(int) + + +class _Sink(ez.Unit): + INPUT = ez.InputStream(int) + + +class _TopicPassthrough(ez.Collection): + IN = ez.InputTopic(int) + OUT = ez.OutputTopic(int) + + def network(self) -> ez.NetworkDefinition: + return ((self.IN, self.OUT),) + + +class _RelayInputPassthrough(ez.Collection): + IN = ez.InputRelay(int, leaky=False, max_queue=None, copy_on_forward=True) + OUT = ez.OutputTopic(int) + + def configure(self) -> None: + self.IN.leaky = True + self.IN.max_queue = 7 + + def network(self) -> ez.NetworkDefinition: + return ((self.IN, self.OUT),) + + +class _RelayOutputPassthrough(ez.Collection): + IN = ez.InputTopic(int) + OUT = ez.OutputRelay(int, num_buffers=16, force_tcp=True, copy_on_forward=False) + + def configure(self) -> None: + self.OUT.num_buffers = 8 + + def network(self) -> ez.NetworkDefinition: + return ((self.IN, self.OUT),) + + +class _TopicSystem(ez.Collection): + SOURCE = _Source() + PASSTHROUGH = _TopicPassthrough() + SINK = _Sink() + + def network(self) -> ez.NetworkDefinition: + return ( + (self.SOURCE.OUTPUT, self.PASSTHROUGH.IN), + (self.PASSTHROUGH.OUT, self.SINK.INPUT), + ) + + +class _InputRelaySystem(ez.Collection): + SOURCE = _Source() + PASSTHROUGH = _RelayInputPassthrough() + SINK = _Sink() + + def network(self) -> ez.NetworkDefinition: + return ( + (self.SOURCE.OUTPUT, self.PASSTHROUGH.IN), + (self.PASSTHROUGH.OUT, self.SINK.INPUT), + ) + + +class _OutputRelaySystem(ez.Collection): + SOURCE = _Source() + PASSTHROUGH = _RelayOutputPassthrough() + SINK = _Sink() + + def network(self) -> ez.NetworkDefinition: + return ( + (self.SOURCE.OUTPUT, self.PASSTHROUGH.IN), + (self.PASSTHROUGH.OUT, self.SINK.INPUT), + ) + + +def test_input_output_topics_behave_as_shortcuts(): + system = _TopicSystem() + ctx = ExecutionContext.setup({"SYSTEM": system}) + assert ctx is not None + assert (system.SOURCE.OUTPUT.address, system.PASSTHROUGH.IN.address) in ctx.connections + assert (system.PASSTHROUGH.IN.address, system.PASSTHROUGH.OUT.address) in ctx.connections + assert (system.PASSTHROUGH.OUT.address, system.SINK.INPUT.address) in ctx.connections + + +def test_input_relay_rewrites_edges_and_syncs_settings(): + system = _InputRelaySystem() + ctx = ExecutionContext.setup({"SYSTEM": system}) + assert ctx is not None + + relay = system.PASSTHROUGH.components["__relay_in_IN"] + source = system.SOURCE.OUTPUT.address + endpoint_in = system.PASSTHROUGH.IN.address + endpoint_out = system.PASSTHROUGH.OUT.address + sink = system.SINK.INPUT.address + + assert (source, endpoint_in) in ctx.connections + assert (endpoint_in, relay.INPUT.address) in ctx.connections + assert (relay.OUTPUT.address, endpoint_out) in ctx.connections + assert (endpoint_out, sink) in ctx.connections + assert (endpoint_in, endpoint_out) not in ctx.connections + + assert relay.SETTINGS.leaky is True + assert relay.SETTINGS.max_queue == 7 + assert relay.SETTINGS.copy_on_forward is True + + +def test_output_relay_rewrites_edges_and_syncs_settings(): + system = _OutputRelaySystem() + ctx = ExecutionContext.setup({"SYSTEM": system}) + assert ctx is not None + + relay = system.PASSTHROUGH.components["__relay_out_OUT"] + source = system.SOURCE.OUTPUT.address + endpoint_in = system.PASSTHROUGH.IN.address + endpoint_out = system.PASSTHROUGH.OUT.address + sink = system.SINK.INPUT.address + + assert (source, endpoint_in) in ctx.connections + assert (endpoint_in, relay.INPUT.address) in ctx.connections + assert (relay.OUTPUT.address, endpoint_out) in ctx.connections + assert (endpoint_out, sink) in ctx.connections + + assert relay.SETTINGS.num_buffers == 8 + assert relay.SETTINGS.force_tcp is True + assert relay.SETTINGS.copy_on_forward is False + + +def test_metadata_separates_collection_topics_relays_and_unit_streams(): + system = _InputRelaySystem() + ctx = ExecutionContext.setup({"SYSTEM": system}) + assert ctx is not None + + runner = ez.GraphRunner(components={"SYSTEM": system}) + metadata = runner._component_metadata() + + passthrough_meta = metadata.components[system.PASSTHROUGH.address] + assert isinstance(passthrough_meta, CollectionMetadata) + assert "IN" in passthrough_meta.relays + assert isinstance(passthrough_meta.relays["IN"], InputRelayMetadata) + assert passthrough_meta.relays["IN"].leaky is True + assert passthrough_meta.relays["IN"].max_queue == 7 + assert "OUT" in passthrough_meta.topics + assert isinstance(passthrough_meta.topics["OUT"], OutputTopicMetadata) + + source_meta = metadata.components[system.SOURCE.address] + sink_meta = metadata.components[system.SINK.address] + assert isinstance(source_meta, UnitMetadata) + assert isinstance(source_meta.streams["OUTPUT"], OutputStreamMetadata) + assert isinstance(sink_meta, UnitMetadata) + assert isinstance(sink_meta.streams["INPUT"], InputStreamMetadata) diff --git a/tests/test_topology_api.py b/tests/test_topology_api.py new file mode 100644 index 00000000..4633c093 --- /dev/null +++ b/tests/test_topology_api.py @@ -0,0 +1,105 @@ +import asyncio + +import pytest + +from ezmsg.core.graphcontext import GraphContext +from ezmsg.core.graphmeta import TopologyChangedEvent, TopologyEventType +from ezmsg.core.graphserver import GraphService +from ezmsg.core.processclient import ProcessControlClient + + +async def _next_matching_event( + stream, predicate, timeout: float = 1.0 +) -> TopologyChangedEvent: + async def _wait() -> TopologyChangedEvent: + while True: + event = await anext(stream) + if predicate(event): + return event + + return await asyncio.wait_for(_wait(), timeout=timeout) + + +@pytest.mark.asyncio +async def test_topology_subscription_reports_session_edge_changes(): + graph_server = GraphService().create_server() + address = graph_server.address + + owner = GraphContext(address, auto_start=False) + observer = GraphContext(address, auto_start=False) + + await owner.__aenter__() + await observer.__aenter__() + + stream = observer.subscribe_topology_events(after_seq=0) + try: + await owner.connect("SRC", "DST") + event = await _next_matching_event( + stream, + lambda e: ( + e.event_type == TopologyEventType.GRAPH_CHANGED + and "DST" in e.changed_topics + ), + timeout=1.0, + ) + assert event.source_session_id == str(owner._session_id) + + await owner.disconnect("SRC", "DST") + event = await _next_matching_event( + stream, + lambda e: ( + e.event_type == TopologyEventType.GRAPH_CHANGED + and "DST" in e.changed_topics + ), + timeout=1.0, + ) + assert event.source_session_id == str(owner._session_id) + finally: + await stream.aclose() + await owner.__aexit__(None, None, None) + await observer.__aexit__(None, None, None) + graph_server.stop() + + +@pytest.mark.asyncio +async def test_topology_subscription_reports_process_changes(): + graph_server = GraphService().create_server() + address = graph_server.address + + observer = GraphContext(address, auto_start=False) + await observer.__aenter__() + + process = ProcessControlClient(address) + stream = observer.subscribe_topology_events(after_seq=0) + + try: + await process.connect() + assert process.client_id is not None + process_key = process.client_id + await process.register(["SYS/U1"]) + + registered = await _next_matching_event( + stream, + lambda e: ( + e.event_type == TopologyEventType.PROCESS_CHANGED + and e.source_process_id == process_key + ), + timeout=1.0, + ) + assert registered.source_session_id is None + + await process.update_ownership(added_units=["SYS/U2"], removed_units=["SYS/U1"]) + updated = await _next_matching_event( + stream, + lambda e: ( + e.event_type == TopologyEventType.PROCESS_CHANGED + and e.source_process_id == process_key + ), + timeout=1.0, + ) + assert updated.source_session_id is None + finally: + await stream.aclose() + await process.close() + await observer.__aexit__(None, None, None) + graph_server.stop()