fix(desktop): take the macOS queue off the Desktop Tests critical path - #95852
fix(desktop): take the macOS queue off the Desktop Tests critical path#95852posthog[bot] wants to merge 1 commit into
Conversation
The p95 of Desktop Tests is queue time for a depot-macos-26 slot, not test time: the macOS job runs in a flat 5m30s on every sampled run, while its wait for a slot ranges from one second to sixty minutes. Move the web E2E suite to its own Linux job. It drives a plain SPA in stock Chromium, so it needs no macOS and no longer holds a slot in the fixed pool. Drop integration-test from the e2e job's needs. That Linux job only requires a green unit run, so waiting on macOS put the whole pool queue on its critical path. With one suite left in the macOS job, its continue-on-error handling and the trailing verdict step are no longer needed. Generated-By: PostHog Desktop Task-Id: ddbc4e41-e31d-41b3-9504-4fe96d04fa9b
|
Merging to
After your PR is submitted to the merge queue, this comment will be automatically updated with its status. If the PR fails, failure details will also be posted here |
🦔 PostHog Review reviewed this pull requestNothing worth raising this time, so here's a calming picture instead:
|
🤖 CI report
|
![stamphog[bot]](https://avatars.githubusercontent.com/in/3000332?s=60&v=4){kind=small}
There was a problem hiding this comment.
Not approved yet — waiting on the conditions below.
Re-add the stamphog label to request another review once you have addressed this.
Gates denied: this is a CI workflow change (deny-listed infra/CI territory) classified as never-auto-approvable, and it has zero human or agent reviews providing independent assurance.
- Modifies .github/workflows/desktop-test.yml, which is on the CI/infra deny-list and gated as T2-never — requires human review regardless of diff content.
- No reviews or approvals present to serve as independent assurance for this risky-territory change.
Gate mechanics and policy version
| Gate | Result | |
|---|---|---|
| prerequisites | ✓ | all clear |
| deny-list | ✗ | matches: infra_cicd |
| size | ✓ | 92L, 1F substantive — within ceiling |
| tier | ✗ | classified as T2-never: T2-never (92L, 1F, single-area, fix) |
| stamphog 2.0.0b4 | .stamphog/policy.yml @ 651701c · reviewed head 651701c |
Problem
depot-macos-26slot is what moves, from one second to sixty minutes.depot-macos-26is a fixed pool with FIFO queuing, and five desktop workflows share it.e2ejob waited on the macOS job, so the whole macOS queue sat on its critical path.Measured on the
integration-testjob, over sampled successful runs either side of 2026-09-03:The suspected cold turbo cache is not the cause
The report suspected that dda7614 left
build:depsrunning cold on macOS. Per-step timings say otherwise.Build packagestook 8-13s before that commit and 6-13s after, so it never went cold. Depot Cache serves turbo on these runners, which is what the commit assumed. No cache is restored here.Changes
e2ejob now starts as soon as the unit job passes, instead of queueing behind macOS. A red tree still never reaches the gateway.continue-on-errorhandling and the trailing verdict step are gone. Mechanical.Desktop Tests Passgates on the newweb-testjob. Mechanical.Before:
flowchart LR changes --> unit-test changes --> integration-test unit-test --> e2e integration-test --> e2e integration-test -. holds a macOS slot .-> webe2e[web E2E suite] e2e --> pass[Desktop Tests Pass] integration-test --> pass style integration-test fill:#F54E00,color:#fff style webe2e fill:#F54E00,color:#fffAfter:
flowchart LR changes --> unit-test changes --> integration-test changes --> web-test unit-test --> e2e e2e --> pass[Desktop Tests Pass] integration-test --> pass web-test --> pass style integration-test fill:#F54E00,color:#fff style web-test fill:#1D4AFF,color:#fffRelated but separate: #94509 reduces macOS pool pressure from the release workflow. This PR touches only
desktop-test.yml.How did you test this code?
pnpm run test:e2e:webwithCI=true, 2 passed. The job's build step (turbo build --filter=@posthog/web^...) and Chromium install were run the same way first.products/desktop/apps/web/playwright-report/.actionlinton the changed workflow, with the repo's config and the parallel-steps flattener: clean.hogli lint:workflows: 9 of 9 checks pass across 133 workflows.Automatic notifications
Docs update
None.
🤖 Agent context
Autonomy: Fully autonomous
/writing-simplified-technical-english, plus the repo'sdepot-github-runnersskill for the fixed-pool and Depot Cache behavior.concurrencygroup so master pushes stop serializing. That would cancel per-commit test signal on master, which is a call for the desktop team rather than a latency fix.gh pr list --state open --searchon the workflow name and on the symptom found no PR touchingdesktop-test.yml.Created with PostHog Desktop from this inbox report.