Repository navigation
[connectors] Run the Delta CDC suspend case on aarch64 again - #7969
Conversation
`cdc_file_suspend_test` has been skipped on aarch64 since June 2025, when it failed intermittently right after the arm64 adapter tests moved to self-hosted runners that ran them with no concurrency limit; the case then waited 30 s per chunk. CI now caps the number of concurrent tests, and the case waits up to 600 s per chunk. Fixes #4176. Signed-off-by: Leonid Ryzhyk <ryzhyk@gmail.com>
| ignore_if(target_arch = "aarch64", "flaky on aarch64"), | ||
| requires(Storage, Checkpoints, SuspendResume) | ||
| )] | ||
| #[engine_test(requires(Storage, Checkpoints, SuspendResume))] |
There was a problem hiding this comment.
pitfall (medium): The 600 s argument only covers wait_for_records_materialized. This case also runs the suspend path in test_cdc, which still waits on a hardcoded timeout(Duration::from_secs(100), receiver.recv()), once per chunk (11 suspends). That is the only wait that only this case reaches, so it is the one an arm64 runner under load would hit first. The June 2025 log has expired, so we can't tell which wait failed. Use Duration::from_millis(default_timeout_ms()) there too, so that the reasoning in the commit message holds for every wait in the case.
|
What I ran (x86_64, debug; no arm64 host was available):
Checked the claims: Flaky-test gate, medium: this passes provided the 100 s suspend timeout also moves to the default wait (see the inline comment). The commit message gives a plausible cause but no evidence for it, and the original failure log has expired. This PR's own aarch64 adapter job is the only real evidence that the case is stable there. Please re-run that job a few times before merging. If the case flakes again, reopen #4176 instead of skipping it again without an issue. |
cdc_file_suspend_testhas been skipped on aarch64 since June 2025, when it failed intermittently right after the arm64 adapter tests moved to self-hosted runners that ran them with no concurrency limit; the case then waited 30 s per chunk. CI now caps the number of concurrent tests, and the case waits up to 600 s per chunk.Linked Issue
Fixes #4176.
Describe Manual Test Plan
Checklist
Breaking Changes?
Mark if you think the answer is yes for any of these components:
Describe Incompatible Changes