mirror of
https://github.com/coder/coder.git
synced 2026-09-21 20:51:01 +08:00
Bobby Ho
bd693ad4ae
`test-go-pg (macos-latest)` has recurred three times on [coder/internal#1365](https://github.com/coder/internal/issues/1365), most recently as the `cli` package hitting its own `go test -timeout 20m` after several minutes of runner-side scheduling delay, which marks hundreds of in-flight subtests as failed instead of surfacing the actual cause. Raise `GOTEST_FLAGS -timeout` from `20m` to `25m` and the matching `timeout-minutes` on every job that shares it (`test-go-pg`, `test-go-pg-17`, `test-go-race-pg` in `ci.yaml`; `flake_go` in `flake-go.yaml`; `test-go-pg` in `nightly-gauntlet.yaml`) from `25` to `30`, preserving the 5-minute margin that lets `go test`'s own timeout fire before GitHub cancels the job and produce a goroutine dump for a genuine hang. Also lower macOS `test-parallelism-tests` from `16` to `8`, matching Linux, to test whether the 2x core oversubscription is contributing scheduler contention. Windows keeps `16`; nothing in the current evidence implicates it. This is a mitigation, not a confirmed root cause; the runner scheduling delay itself is still unexplained. Refs coder/internal#1365 ## AI disclosure Per [AI Contribution guidelines](https://coder.com/docs/about/contributing/AI_CONTRIBUTING): this PR was primarily authored by Claude Code, directed by @BobbyHo, based on triage of the linked ticket's CI-flake-bot comments and a review of the current `Makefile`/`ci.yaml`/`flake-go.yaml`/`nightly-gauntlet.yaml` timeout configuration. **Verification:** - `make lint/actions/actionlint` and `make lint/actions/zizmor` both pass clean on the three edited workflow files. - `make pre-commit` passes (includes `make gen`, `make fmt`, `make lint`, and a full local build). - All three edited workflow YAMLs parse (`ruby -ryaml`) and `make -n test TEST_PACKAGES="./cli/..."` confirms `GOTEST_FLAGS` resolves to `-timeout 25m` as intended. - This is a config-only change (no application code touched); the intended effect can only be confirmed by watching real `test-go-pg` (macOS) runs on this PR and on `main` after merge, since the underlying runner-stall failure is intermittent.