[
https://issues.apache.org/jira/browse/CASSANDRA-20829?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18013644#comment-18013644
]
Branimir Lambov commented on CASSANDRA-20829:
---------------------------------------------
How about adding a new method in the secondary index interface, so that we
don't do this when we know we don't need to?
I'm not sure going through the compaction path will work well for 4.1, because
we might find overlaps and leave some tombstones in, especially if things like
unsafe aggressive expiration are in use (which we often have to turn on for a
while to solve a problem). Those tombstones would be placed in random buckets,
which will make things worse.
My opinion is that it would be better to implement the expiration as a separate
step, during the initialization of the compaction task if necessary (and
preferably separately for trunk).
> Secondary index implementations do not integrate with IndexGCTransaction when
> compaction contains fully expired SSTables
> ------------------------------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-20829
> URL: https://issues.apache.org/jira/browse/CASSANDRA-20829
> Project: Apache Cassandra
> Issue Type: Bug
> Components: Feature/2i Index, Local/Compaction, Local/Compaction/TWCS
> Reporter: Stefan Miklosovic
> Assignee: Stefan Miklosovic
> Priority: Normal
> Fix For: 4.0.x, 4.1.x
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> There is a test (1) which ensures that when data are TTLed and compacted,
> IndexGCTransaction is aware of that and it will invoke Indexer.removeRow()
> method eventually.
> However, this is not working properly when we have fully expired SSTables,
> e.g. as the result of a table being on TWCS and having TTL on that.
> The reason is that in CompactionTask, we are filtering out fully expired ones
> (2). These then do not go to the compaction process and then they are not
> reacted on in listener() (3) which contains this logic (4). Eventually,
> onRowMerge in IndexGCTransaction will make the diff and in its commit
> indexer.removeRow(row); will notify 2i about its removal.
>
> This integration is missing and it is quite a big problem because if there
> are custom secondary index implementations the fact that SSTables were fully
> expired is not propagated to them which means that data are never removed
> from whatever backend they use.
> The solution is to go to the compaction with fully expired SSTables as well
> _but only if we detected that respective column family has some indexes_
>
> (1)
> [https://github.com/apache/cassandra/blob/cassandra-4.1/test/unit/org/apache/cassandra/index/CustomIndexTest.java#L583-L607]
> (2)
> [https://github.com/apache/cassandra/blob/cassandra-4.1/src/java/org/apache/cassandra/db/compaction/CompactionTask.java#L174]
> (3)
> [https://github.com/apache/cassandra/blob/cassandra-4.1/src/java/org/apache/cassandra/db/compaction/CompactionIterator.java#L130]
> (4)
> [https://github.com/apache/cassandra/blob/cassandra-4.1/src/java/org/apache/cassandra/db/compaction/CompactionIterator.java#L235-L252]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
