Perf: add in-process timing instrumentation to replace py-spy dependency

[alxmrs]

Problem

The current profiling workflow in perf_tests/ requires py-spy, which needs sudo on macOS:

sudo ./profile.sh groupby_air_full.py

This makes profiling unavailable in many CI environments and for users without root access. Additionally, cProfile only sees the Python side of query execution — the entire DataFusion execution appears as a single opaque to_arrow_table() call (0.66s for 320 MB of data in our benchmarks), making it impossible to distinguish:

• Time spent in xarray I/O (ds.isel(block), .values)
• Time spent in numpy → Arrow conversion (iter_record_batches)
• Time spent in DataFusion (aggregation, hash-grouping, row filtering)
• Time spent waiting for GIL between partitions

Proposed solution

Add a lightweight per-partition timing mode to the _iteration_callback mechanism already in read_xarray_table:

# reader.py — extend make_stream to record timing
def make_stream(projected_names=None) -> pa.RecordBatchReader:
    if _timing_callback is not None:
        t0 = time.perf_counter()
    result = pa.RecordBatchReader.from_batches(...)
    if _timing_callback is not None:
        _timing_callback({"elapsed_s": time.perf_counter() - t0, "block": block})
    return result

Or alternatively: add a verbose=True flag to read_xarray_table that prints per-partition timings to stderr, giving insight into whether execution is I/O-bound or compute-bound without requiring external tools.

A lower-overhead option is to expose a PartitionTimingHook via LazyArrowStreamTable that fires before/after each factory call.

Current profiling workaround

For the query_large_synthetic.py benchmark (320 MB, 8 partitions):

Full scan AVG:    0.662s  (341 MB peak)   ← all in Rust's to_arrow_table()
Filter pushdown:  0.005s  (0.7 MB peak)   ← 7/8 partitions pruned
GROUP BY lat/lon: 0.839s  (341 MB peak)   ← +27% vs. AVG for hash aggregation

cProfile shows 246 Python calls for the full scan, but 0.660s of those are attributed to {method 'to_arrow_table' of 'datafusion.DataFrame' objects} — a Rust builtin that cProfile cannot introspect.

Related

• _iteration_callback in XarrayRecordBatchReader and read_xarray_table — already used for testing
• perf_tests/profile.sh — current py-spy based approach
• perf_tests/query_large_synthetic.py — new benchmark script added alongside this issue