mirror of
https://github.com/cozystack/cozystack.git
synced 2026-11-04 02:57:27 +09:00
myasnikovdaniil
aa707f17b5
## What this PR does The etcd suite's metrics contract failed in CI with `metrics endpoint did not return 200 over http on the 'metrics' port; got: If you don't see a command prompt, try pressing enter.` The probe runs a one-shot curl Pod and reads a `code=<status>` marker back off its output via `kubectl run --rm --restart=Never --attach`. `--attach` only establishes its connection once the Pod reaches Running, so a container that exits in milliseconds — which is what curl against a Pod IP does — writes its stdout into the void, and `--rm` then destroys the Pod carrying the log that still held it. The step reports the metrics endpoint as broken on a cluster that was serving it correctly throughout. The evidence that this was output loss rather than an endpoint failure is that the captured output contained no `code=` at all. A genuine connection failure would still have printed `code=000` from `-w`, plus curl's own error text from `-fsS`, and both were being captured. Meanwhile the probe Pod's events show a clean run, all three etcd members were up without restarts, and every preceding assert in the suite passed. The fix drops `--rm`, re-reads the Pod log when the attach stream carried no marker, and deletes the Pod explicitly afterwards. This is the same treatment `hack/e2e-chainsaw/_lib/talos-image-cache.sh` already applies to the identical race — that probe documents it inline; the etcd one was simply never given it. A `--pod-running-timeout` is set for the same reason it is there: the probe image is used nowhere else in the tree, so it pays a cold pull on every run and kubectl's one-minute default was a second way to end up with no marker. The recovery cannot turn a real failure green. The marker encodes the outcome, so a `code=503` recovered from the log still fails the step; an attach stream that already carried a marker is never second-guessed; and the pre-run delete blocks, so a stale Pod's log cannot be read in place of this run's. Verified by running the probe's exact curl invocation against a 503 and against a dead port: the first yields `code=503`, the second `code=000` plus curl's error, and both fail. The probe still runs exactly once — this is output recovery, not a retry. Helper extracted to `hack/e2e-chainsaw/_lib/etcd-probe.sh` with unit coverage in `hack/etcd-probe_test.bats`. ### Release note ```release-note NONE ``` <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Bug Fixes** * Improved etcd metrics probe reliability by recovering probe results when attached output is incomplete. * Preserved meaningful probe and Kubernetes error details, including non-success status codes. * Improved probe pod timing and cleanup to reduce intermittent end-to-end test failures. * **Tests** * Added coverage for successful recovery, failed probes, missing output, and log retrieval errors. <!-- end of auto-generated comment: release notes by coderabbit.ai -->