Martijn Visser created FLINK-41455:
--------------------------------------
Summary:
TaskExecutorPartitionLifecycleTest.testEnableBatchJobRecoveryAndNotRetainPartitions
depends on the RPC timer firing late
Key: FLINK-41455
URL: https://issues.apache.org/jira/browse/FLINK-41455
Project: Flink
Issue Type: Technical Debt
Components: Runtime / RPC
Reporter: Martijn Visser
The test asserts that the partition cleanup future does not complete within 15
seconds. TaskExecutor#disconnectJobManagerConnection schedules that cleanup
with scheduleRunAsync and the same 15 second delay, and it does so before the
test starts waiting. The test therefore only passes when the timer fires later
than the test's wait ends.
With the Pekko RPC service this holds: 60 timers of 15 seconds fired between
7.5 and 21.5 ms late (median 16.9 ms), and the test passed 100 out of 100 runs.
With an RPC service whose timers fired 0.2 to 10.9 ms late, the same test
failed 14 out of 100 runs.
It does not fail on master today. It would start failing with any change that
makes scheduled RPC calls more punctual. The test should not wait for the same
duration as the delay it checks.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
