Github user StephanEwen commented on the issue:
https://github.com/apache/flink/pull/5601
S3 is actually strongly consistent when reading newly created objects, just
not in listing or renaming objects (files).
The test seems to actually use reads of full paths, so wondering why there
is a failure in the first place.
If there is an issue that the Yarn upload code relies on eventually
consistent operations, then fixing the test by retries may disguise the actual
issue. If there is no eventually consistent operation, then this should not be
necessary in the first place. I fear this change may be down a tricky path...
Can you explain/double check why the failure happened and why the retry is
necessary to stabilize the test, but the actual Yarn code is not affected by
this?
---
