Fix and port Memory Pressure Events feature (#865) by erikdarlingdata · Pull Request #866 · erikdarlingdata/PerformanceMonitor

erikdarlingdata · 2026-04-19T19:18:53Z

Summary

Fixes the Memory Pressure Events chart in Dashboard and ports the full feature to Lite, which was missing it entirely. Fixes #865.

Dashboard fix

Chart previously filtered to HIGH severity only (indicator >= 3). On most servers that threshold is essentially never crossed, so the chart was empty even when sp_pressuredetector-level medium pressure (indicator = 2) was firing constantly — which is exactly what the reporter on #865 was seeing.

Switched to stacked bar chart: count of pressure events per hour, split by SQL Server (process) vs Operating System (system), with severe (indicator >= 3) bars stacked on top of medium (indicator = 2) bars in a darker shade.
Relabeled series from opaque "Process" / "System" to "SQL Server" / "Operating System".
Extended ChartHoverHelper to support BarPlot so tooltips work; integer counts format without trailing .0.
Added MCP interpretation guidance: the get_memory_pressure_events tool description now explains the indicator scale, and McpInstructions.cs has an Interpreting Memory Pressure Events section with process-vs-system interpretation, follow-up-tool tables, and named patterns (bursty, flat-line, only-one-side-firing, etc.).

Lite port

Lite had no memory pressure events feature at all. This adds the full stack:

Schema: new memory_pressure_events table + index, DuckDB schema v25, added to ArchivableTables, server-id-fix list, and ArchiveService.
Collector: CollectMemoryPressureEventsAsync queries RING_BUFFER_RESOURCE_MONITOR and client-side-dedupes against DuckDB's MAX(sample_time). Azure SQL DB returns zero rows (ring buffer not exposed there). Scheduled every 5 min in Aggressive/Balanced presets, 15 min in Low-Impact.
UI: new "Memory Pressure Events" sub-tab on the Memory tab with the same stacked-bar chart as Dashboard.
Hover: ported BarPlot support to Lite's ChartHoverHelper.
MCP: new get_memory_pressure_events tool and the same Interpreting Memory Pressure Events guidance section.

Test plan

Dashboard builds clean
Dashboard chart renders pressure events on SQL2025 30-day window; tooltip shows label / count / time
Dashboard legend reads "SQL Server" / "Operating System" with medium/severe shades
Lite builds clean
Lite launches, collector registered, sub-tab visible and chart renders (limited data since collector just started)
Verify on a server with actual severe (indicator >= 3) events that stacked caps render correctly
Confirm "No memory pressure events in selected time range" message renders when a window has no matching rows
Let Lite run for 24h+ and confirm incremental collection keeps adding new events without duplicates

Not changed

Dashboard collector (collect.memory_pressure_events_collector) and its schema unchanged.
Lite's existing memory sub-tabs (Overview, Memory Clerks, Memory Grants) unchanged.

Chart previously filtered to HIGH severity only (indicator>=3), which on most servers never fires, producing an empty chart even when sp_pressuredetector- level medium pressure (indicator=2) was occurring constantly. Switch to stacked bars per hour, split by SQL Server (process) vs Operating System (system), with severe events capped on top of medium in a darker shade. Extend ChartHoverHelper to support BarPlot tooltips. Add MCP guidance for interpreting indicator values and routing to the right follow-up tool. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

erikdarlingdata

Review summary

What this PR does

Fixes #865: Memory Pressure Events chart was filtering on Severity == "HIGH" (effectively indicator >= 3), which almost never fires. New filter is indicator >= 2, matching sp_pressuredetector's medium/severe semantics.
Switches the chart from a scatter line to four stacked bar series (SQL Server medium/severe + OS medium/severe), drawn at symmetric X offsets per hour bucket, so SQL-process vs OS-level pressure are visually separable.
Teaches ChartHoverHelper to handle BarPlot in addition to Scatter, and formats integer-valued tooltips without .0.
Expands the get_memory_pressure_events MCP tool [Description] and adds an "Interpreting Memory Pressure Events" section to McpInstructions.cs — indicator scale, process vs system semantics, follow-up tool table, common patterns.

Repo-policy checks

Base branch: dev ✅
PlanAnalyzer: not touched, no sync concern ✅
Schema / upgrade scripts: no SQL changes ✅
Lite-first: PR description explicitly confirms Lite has neither this chart nor a memory-pressure MCP tool. Verified — no MemoryContent.* under Lite/ and no matching tool in Lite/Mcp/. ✅
SignPath / build.yml: no workflow changes ✅

What's good

Stacked-bar geometry is right: sqlSevereBars uses ValueBase = sqlMedium, Value = sqlMedium + sqlSevere so severe sits on top of medium. FindNearestBar correctly reports the segment height (Value - ValueBase), so the tooltip shows the individual segment's count rather than the running total. That's the non-obvious bit and it's handled.
Y-axis floor of 5.0 avoids a single-event day rendering as one huge bar dominating the plot.
The new MCP section is genuinely useful — indicator scale + the "system-only fires but process does not" guidance is exactly what an LLM needs to avoid blaming SQL Server for host-level pressure.

Needs attention (see inline comments)

ChartHoverHelper.cs:204-206 — the integer-formatting change affects every scatter chart using the helper, not just the new bar chart. Likely cosmetic regression on whole-number samples elsewhere (memory, PLE, wait-time charts). Scope it to bar-originated hovers or make it construction-time configurable.
ChartHoverHelper.cs:134-137 — FindNearestBar does two Plot.GetPixel calls × every bar, every mouse move. Hoist halfWidthPx out of the inner loop. Not a blocker, just a lurking perf issue at 30-day windows.
MemoryPressureEventItem.Severity (Models/MemoryPressureEventItem.cs:13) is now dead — last reader was the filter this PR removed. Drop it or mark it.
MemoryContent.xaml.cs:1227 — maxBarCount takes max(sqlTotal, osTotal) rather than sum; correct only because SQL and OS bars are X-offset rather than stacked. Worth a one-line comment to protect against future "fixes".

Test coverage
No new tests. The indicator-filter and per-hour bucketing logic is pure data transformation and would be trivially testable if extracted from LoadMemoryPressureEventsChart. Not a blocker — Dashboard UI code isn't currently tested — but a natural place to start if you ever want regression coverage on chart inputs.

Task list
Two unchecked items in the PR description (verifying severe-event stacking + the empty-state message on a live server) are the right things to confirm before merge.

Generated by Claude Code

erikdarlingdata · 2026-04-19T19:21:46Z

+                string valueFormatted = (bestPoint.Y == Math.Floor(bestPoint.Y))
+                    ? bestPoint.Y.ToString("N0")
+                    : bestPoint.Y.ToString("N1");


This format switch is cross-cutting: it fires for every ChartHoverHelper consumer, not just the new bar chart. Scatter values that happen to be whole numbers (e.g. a memory trend reading of exactly 100 GB or a PLE of 1000) will now render as 100 / 1000 while neighbouring samples render as 99.7 / 1000.4 — inconsistent unit precision in the same tooltip.

If the goal is integer-looking counts for the bar chart only, guard it on _barPlots having produced the winner (e.g. track bool bestIsBar through FindNearestBar) or parameterise the helper's format at construction time. Otherwise this will likely show up as a cosmetic regression on the other memory/wait/cpu charts.

Generated by Claude Code

erikdarlingdata · 2026-04-19T19:21:52Z

+                var topPixel = _chart.Plot.GetPixel(new ScottPlot.Coordinates(bar.Position, bar.Value));
+                double halfWidthPx = Math.Abs(
+                    _chart.Plot.GetPixel(new ScottPlot.Coordinates(bar.Position + bar.Size / 2, bar.Value)).X
+                    - topPixel.X);


Two Plot.GetPixel calls per bar, per mouse move. A 30-day window at 24h buckets × up to 4 stacked series = up to ~2,880 bars, re-computed every 30 ms while the mouse is inside the chart. halfWidthPx depends only on bar.Size (constant per series here), so hoist it to a per-series calculation outside the inner loop. Also consider pixel-space binary-searching on Position once the bars are sorted — the inner foreach is effectively an O(N) scan per mouse event.

Not a blocker for this PR, but worth a follow-up if the 30-day view feels laggy on hover.

Generated by Claude Code

erikdarlingdata · 2026-04-19T19:21:57Z

+                    int sqlMedium = g.Count(d => d.MemoryIndicatorsProcess == 2);
+                    int sqlSevere = g.Count(d => d.MemoryIndicatorsProcess >= 3);
+                    int osMedium = g.Count(d => d.MemoryIndicatorsSystem == 2);
+                    int osSevere = g.Count(d => d.MemoryIndicatorsSystem >= 3);


After this PR MemoryPressureEventItem.Severity (Models/MemoryPressureEventItem.cs:13) is no longer consumed anywhere in the Dashboard — the last reader was the old d.Severity.Equals("HIGH"…) filter. DatabaseService.Memory.cs:151 still reads column ordinal 6 and populates it, so it's dead-but-still-materialised. Either drop the field and stop selecting the column, or leave a // kept for MCP/grid binding comment so the next person doesn't delete it accidentally.

Generated by Claude Code

erikdarlingdata · 2026-04-19T19:22:02Z

-            var pressureLimits = MemoryPressureEventsChart.Plot.Axes.GetLimits();
-            MemoryPressureEventsChart.Plot.Axes.SetLimitsY(0, pressureLimits.Top * 1.05);
+            MemoryPressureEventsChart.Plot.YLabel("Pressure Events per Hour");
+            MemoryPressureEventsChart.Plot.Axes.SetLimitsY(0, Math.Max(maxBarCount * 1.1, 5.0));


maxBarCount only tracks sqlTotal / osTotal separately (whichever one side is taller). Because SQL and OS bars are drawn at different X offsets (x - barOffset vs x + barOffset) they never stack on top of each other, so this is correct — but it's subtle. Worth a one-line comment next to this so a future reader doesn't "fix" it to sqlTotal + osTotal, which would over-scale the Y axis.

Generated by Claude Code

Lite was missing the RING_BUFFER_RESOURCE_MONITOR collector entirely — no collector, no table, no chart, no MCP tool. This adds the full feature: - Schema: new memory_pressure_events table + index, schema v25, added to ArchivableTables, server-id-fix list, and ArchiveService. - Collector: CollectMemoryPressureEventsAsync queries the ring buffer and client-side-dedupes against DuckDB's MAX(sample_time). Azure SQL DB returns zero rows (ring buffer not exposed there). Scheduled every 5 min (Aggressive and Balanced presets) or 15 min (Low-Impact). - UI: new 'Memory Pressure Events' sub-tab on the Memory tab with the same stacked-bar chart as Dashboard (SQL Server medium/severe, Operating System medium/severe). Wired into full-load and sub-tab-switch refresh paths. - Hover: ported the BarPlot support from Dashboard's ChartHoverHelper so bar tooltips work and report the correct segment height for stacked bars. - MCP: new get_memory_pressure_events tool + the 'Interpreting Memory Pressure Events' guidance section in McpInstructions. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Companion update to the new memory_pressure_events table added in this PR. SchemaStatements_MatchTableCount asserts the total table count; needs to move from 29 to 30 to reflect the new table. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

erikdarlingdata mentioned this pull request Apr 19, 2026

[QUESTION] What should Memory Pressure Events show? #865

Closed

erikdarlingdata commented Apr 19, 2026

View reviewed changes

erikdarlingdata changed the title ~~Fix Memory Pressure Events chart; add MCP interpretation (#865)~~ Fix and port Memory Pressure Events feature (#865) Apr 19, 2026

erikdarlingdata merged commit 641e59f into dev Apr 19, 2026
3 checks passed

erikdarlingdata deleted the fix/865-memory-pressure-chart branch April 19, 2026 20:21

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix and port Memory Pressure Events feature (#865)#866

Fix and port Memory Pressure Events feature (#865)#866
erikdarlingdata merged 3 commits into
devfrom
fix/865-memory-pressure-chart

erikdarlingdata commented Apr 19, 2026 •

edited

Loading

Uh oh!

erikdarlingdata left a comment

Uh oh!

erikdarlingdata Apr 19, 2026

Uh oh!

erikdarlingdata Apr 19, 2026

Uh oh!

erikdarlingdata Apr 19, 2026

Uh oh!

erikdarlingdata Apr 19, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

erikdarlingdata commented Apr 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Dashboard fix

Lite port

Test plan

Not changed

Uh oh!

erikdarlingdata left a comment

Choose a reason for hiding this comment

Review summary

Uh oh!

erikdarlingdata Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

erikdarlingdata Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

erikdarlingdata Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

erikdarlingdata Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

erikdarlingdata commented Apr 19, 2026 •

edited

Loading