Skip to content

Human action 68

Jerry Holland edited this page Nov 14, 2026 · 4 revisions

Windows-only release latency qualification

Canonical work item: #68.

The maintainer selected Windows-only latency qualification for current releases. The matching identified lab-windows-a machine is the qualification host. Linux latency remains unqualified. Functional Windows/POSIX release testing and the separate multi-host workload evidence and budget gates remain required.

Owner role: Repository maintainer and qualification-machine owner. The decision and delivery evidence are on #68; this procedure remains available for releases and any future proposal to add Linux qualification.

Current release check

Prerequisites: The matching Windows machine, a quiet environment, a clean release candidate checkout, Node 24+, Git 2.40+, and the hardware/settings recorded under hosts.lab-windows-a. Do not run competing benchmark or test suites.

  1. Review the candidate commit and baseline provenance. Do not borrow the label for another Windows machine or change engine/session overrides to bypass a mismatch.
  2. Run npm run test:benchmark -- --host lab-windows-a --json.
  3. Verify the reference is hosts.lab-windows-a, latencySkipped is false, and comparisons pass. Retain the candidate, toolchain, and output as issue/CI evidence outside the tracked source tree.
  4. Complete the remaining release gates, including the Windows/POSIX functional matrix. A Windows latency pass supplies no Linux latency or production-budget claim.

On a failure or mismatch, preserve output and investigate. Do not re-record just to clear a regression. Discard measurements taken under load and rerun when quiet. Benchmark fixtures normally remove themselves; inspect any disposable leftovers and retain evidence before cleanup. Never change shared history for qualification.

Adding Linux qualification later

Prerequisites: A real quiet identified Linux machine, a reviewed scope extension, clean candidate checkout, and the supported Node/Git toolchain. A shared hosted runner or Linux container sharing the Windows host cannot qualify latency.

  1. Assign a stable public-safe label and run npm run test:benchmark -- --host <assigned-linux-label> before recording. An unknown identity should skip latency. Inspect all deterministic findings.
  2. Investigate differences against current fixture semantics. Historical Linux publication count 57 predates atomic object retention (#49), so unchanged legacy process counts are not an acceptance condition. Preserve that legacy entry and document the cause of changed counts before recording.
  3. Run npm run benchmark:record -- --host <assigned-linux-label> deliberately, then npm run test:benchmark -- --host <assigned-linux-label> --json.
  4. Review provenance and the baseline diff; preserve other entries and tolerances. Verify the selected identified reference, latencySkipped: false, and passing comparisons on the matching machine. Commit the reviewed baseline with its explanation and update the release scope consistently.

Completion evidence for #68 is the maintainer's Windows-only decision plus the delivered testing contract. This does not complete the separate evidence work in #42.

Publish only public-safe evidence. Redact private paths and credentials.

Back to Human TODO · Benchmark host baselines

Clone this wiki locally