Skip to content

test: harden Stax latency evidence and MCP cancellation - #51

Merged
fcostaoliveira merged 3 commits into
mainfrom
test/stax-latency-correlation
Oct 23, 2026
Merged

fcostaoliveira merged 3 commits into
mainfrom
test/stax-latency-correlation

Conversation

@fcostaoliveira

@fcostaoliveira fcostaoliveira commented Oct 23, 2026 •

Copy link
Copy Markdown
Contributor

Stax wait reports now expose whether scheduler records were observed, compare archived/reported counts, and state that an empty report does not establish no blocking. Older stored documents gain these diagnostics at query time without reimport. This follows native reduced-mode controls that produced CPU data while omitting scheduler records.

The Redis pilot retains individual PING monotonic start/end timestamps and RTT, reports baseline p99 variation, and treats missing latency evidence as inconclusive. New stdio protocol coverage cancels active and queued captures and verifies partial retention, child reaping, borrowed-target survival, no storage and continued responsiveness.

The measured follow-up includes exact paired values, binary/source hashes and isolated diagnostic patches. Five 30-second pairs gave +1,434.11% median paired PING p99 on pinned Stax with 0.34% throughput loss; a main-thread-only prototype gave +80.43% p99 with 0.20% throughput loss and lost thread/off-CPU coverage. Both failed the declared latency budget. Short prototype median passes did not survive longer measurement. No source workaround is selected automatically or adopted as a production fix.

M6 guidance also reproduces the pinned native comparison bug (299 synthetic markers reported as PET samples despite zero CPU samples), records fresh capture identity evidence, and provides explicit extraction commands for its pinned source example. Raw PMU scope remains distinct from per-process counters/CPI.

Validation:

  • Clean macOS full suite: 2,117 passed, 109 skipped; coverage 88.75%, above the earlier baseline.
  • Clean Linux aarch64 full suite with PostgreSQL 16.15: 2,183 passed, 43 skipped.
  • Real pinned Linux Stax + MCP 1.30.0: active/queued cancellation passed over both stdio and streamable HTTP.
  • Live PostgreSQL/MCP acceptance stored a native Linux capture (37 CPU samples, resolved busy function and waits) and imported a real M6 Metal capture (zero CPU samples, 299 spans), retaining recorded M6 labels on the Linux importer; filtering, paging and the TMA resource guard passed.
  • Ruff clean; updated CI pending.

Capture remains experimental. This VM study does not validate other platforms, hardware PMU access or a production overhead guarantee. Privileged M1/M6 Stax CPU acceptance remains pending.

@fcostaoliveira
fcostaoliveira marked this pull request as ready for review October 23, 2026 20:46
@fcostaoliveira
fcostaoliveira merged commit f6b4e7e into main Oct 23, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant