[
https://issues.apache.org/jira/browse/HDDS-7110?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17698141#comment-17698141
]
George Jahad commented on HDDS-7110:
------------------------------------
Fixed in [https://github.com/apache/ozone/pull/3653]
Closing
> Race Condition in TestKeyManagerImpl::testListStatusWithTableCacheRecursive()
> -----------------------------------------------------------------------------
>
> Key: HDDS-7110
> URL: https://issues.apache.org/jira/browse/HDDS-7110
> Project: Apache Ozone
> Issue Type: Bug
> Reporter: George Jahad
> Assignee: George Jahad
> Priority: Minor
>
> There is a race condition this test:
> TestKeyManagerImpl::testListStatusWithTableCacheRecursive(). The test
> depends on the timing of the OM double buffer flush. If the flush happens
> quickly enough, the test passes.
>
> *The Problem:*
> The purpose of the test is to confirm that KeyManagerImpl::listStatus()
> correctly handles a mix of keytable cache and db entries.
> The test uses the "writeClient.createDirectory()" method to create some
> directories. That method uses the normal rpc client protocol request path,
> which writes to the cache and the double buffer. Then
> testListStatusWithTableCacheRecursive() writes some keys to those directories
> using addKeyToTable() which bypasses the cache.
> Then it runs keyManager.listStatus(). If the flush hasn't yet occurred, the
> directories are still in the cache, while the keys that belong to those
> directories are in the keytable. The listStatus() algorithm can't handle
> that inconsistency. It expects that directories are added to the table
> before any elements of those directories.
> This race was actually introduced by me 6 months ago in this PR:
> "HDDS-3231. Cleanup KeyManagerImpl"
>
> *The Fix:*
> I've confirmed this is the problem by disabling the flush. Then the problem
> happens consistently, and I can step through it with the debugger.
> The fix is to bypass the double buffer, (as the test did originally,) by
> writing the directories straight to the table using addKeyToTable, instead of
> writeClient.createDirectory(). I have confirmed the fix and will create a PR
> shortly.
> For some reason, this race condition, (which is in master,) happens more
> frequently in the snapshotting branch, which is how I noticed it.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
