KWOK is advisory (not in gate), but 124 jobs/PR, 6/8 runs failing, slowest cells hit the 15m timeout.
Measured 15-min flux-git cell (job 87189194361): tool install ~4.0m; make build cold (kwok-test/action.yml:82 cache:false) ~2.5m; cluster+registry+flux+gitea ~2.7m; flux reconcile wait ~5.6m; validate+cleanup ~0.5m. Fast helm/argocd cells run 1.5–5m (no flux wait).
Ranked fixes:
- Build aicr once, share via artifact (cells download-artifact instead of
make build) — ~2.5m off ~123 cells, drops goreleaser from per-cell install.
- Flip
setup-go cache:false → true.
- Force flux reconcile (
flux reconcile … --with-source) instead of polling — up to ~5m off the 2 flux deployers; revisit poll cadence in kwok/scripts/validate-scheduling.sh + the chainsaw assert.
- Cut recipe×deployer cross-product to ~N+6 instead of N×6 (deployer path is recipe-independent plumbing).
- Pre-pull/cache the kind node image.
Also: root-cause the uniform Tier-1 failures (all deployers failing together = infra/tooling break, not recipe drift); make advisory status explicit in the check name/summary.
KWOK is advisory (not in
gate), but 124 jobs/PR, 6/8 runs failing, slowest cells hit the 15m timeout.Measured 15-min flux-git cell (job 87189194361): tool install ~4.0m;
make buildcold (kwok-test/action.yml:82cache:false) ~2.5m; cluster+registry+flux+gitea ~2.7m; flux reconcile wait ~5.6m; validate+cleanup ~0.5m. Fast helm/argocd cells run 1.5–5m (no flux wait).Ranked fixes:
make build) — ~2.5m off ~123 cells, drops goreleaser from per-cell install.setup-go cache:false → true.flux reconcile … --with-source) instead of polling — up to ~5m off the 2 flux deployers; revisit poll cadence in kwok/scripts/validate-scheduling.sh + the chainsaw assert.Also: root-cause the uniform Tier-1 failures (all deployers failing together = infra/tooling break, not recipe drift); make advisory status explicit in the check name/summary.