Description
The LearnerSnapshotThrottler will only allow 2 concurrent snapshots to be taken, and if there are already 2 snapshots in progress it will wait up to 200ms for one to complete. This isn't enough time for testHighContentionWithTimeout to consistently pass - on a cold JVM running just the one test I was able to get it to fail 3 times in around 50 runs. This 200ms timeout will be hit if there is a delay between a thread calling LearnerSnapshot snap = throttler.beginSnapshot(false); and throttler.endSnapshot();.
This also erroneously fails on the build server, see https://builds.apache.org/job/PreCommit-ZOOKEEPER-Build/2747/testReport/org.apache.zookeeper.server.quorum/LearnerSnapshotThrottlerTest/testHighContentionWithTimeout/ for an example.
I have bumped the timeout up to 5 seconds (which should be more than enough for warmup / gc pauses), as well as added logging to the catch (Exception e) block to assist in debugging any future issues.
An alternate approach would be to separate out results gathered from the threads, because although we only record true/false there are really three outcomes:
1. The snapshotNumber was <= 2, meaning the individual call operated correctly
2. The snapshotNumber was > 2, meaning the test should definitely fail
3. We were unable to snapshot in the time given, so we can't determine if we should fail or pass (although if we have "enough" successes from #1 with no failures from #2 maybe we would pass the test anyway).
Bumping up the timeout is easier.
Attachments
Attachments
Issue Links
- is related to
-
ZOOKEEPER-3047 flaky test LearnerSnapshotThrottlerTest
- Open