You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@cassandra.apache.org by "Jordan West (Jira)" <ji...@apache.org> on 2020/05/04 19:55:00 UTC
[jira] [Commented] (CASSANDRA-14248) SSTableIndex should not use
Ref#globalCount() to determine when to delete index file
[ https://issues.apache.org/jira/browse/CASSANDRA-14248?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17099300#comment-17099300 ]
Jordan West commented on CASSANDRA-14248:
-----------------------------------------
[~jbellis] thanks for picking this up. I couldn't access the link you shared. Looks like a typo. Its been a bit since I filed this so I need to page it back in but quickly, I think the issue in {{SSTableTidier}} is {{SSTable.discoverComponentsFor(...)}} does not actually find the index files ([https://github.com/apache/cassandra/blob/trunk/src/java/org/apache/cassandra/db/lifecycle/LogTransaction.java#L356)|https://github.com/apache/cassandra/blob/trunk/src/java/org/apache/cassandra/db/lifecycle/LogTransaction.java#L356);]. Meanwhile, the check of whether or not `globalCount() == 0` won't always be true when the index is released for the last time (e.g. when removing an index).
It would be good to write some tests to validate my guesses above.
> SSTableIndex should not use Ref#globalCount() to determine when to delete index file
> ------------------------------------------------------------------------------------
>
> Key: CASSANDRA-14248
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14248
> Project: Cassandra
> Issue Type: Bug
> Components: Feature/SASI
> Reporter: Jordan West
> Assignee: Jordan West
> Priority: Normal
> Fix For: 3.11.x
>
>
> {{SSTableIndex}} instances maintain a {{Ref}} to the underlying {{SSTableReader}} instance. When determining whether or not to delete the file after the last {{SSTableIndex}} reference is released, the implementation uses {{sstableRef.globalCount()}}: [https://github.com/apache/cassandra/blob/trunk/src/java/org/apache/cassandra/index/sasi/SSTableIndex.java#L135.] This is incorrect because {{sstableRef.globalCount()}} returns the number of references to the specific instance of {{SSTableReader}}. However, in cases like index summary redistribution, there can be more than one instance of {{SSTableReader}}. Further, since the reader is shared across multiple indexes, not all indexes see the count go to 0. This can lead to cases where the {{SSTableIndex}} file is incorrectly deleted or not deleted when it should be.
>
> A more correct implementation would be to either:
> * Tie into the existing {{SSTableTidier}}. SASI indexes already are SSTable components but are not cleaned up by the {{SSTableTidier}} because they are not found with the currently cleanup implementation
> * Revamp {{SSTableIndex}} reference counting to use {{Ref}} and implement a new tidier.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@cassandra.apache.org
For additional commands, e-mail: commits-help@cassandra.apache.org