[ 
https://issues.apache.org/jira/browse/CASSANDRA-5342?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13701217#comment-13701217
 ] 

Jonathan Ellis commented on CASSANDRA-5342:
-------------------------------------------

LGTM.

Nit: since we only ever use either the metadata, or the anscestors, it may be 
cleaner to split deserializeAncestors into a separate method.  OTOH that might 
make it easier to make mistakes in compatibility.  I'm fine either way.
                
> ancestors are not cleared in SSTableMetadata after compactions are done and 
> old SSTables are removed
> ----------------------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-5342
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-5342
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>    Affects Versions: 1.1.3
>            Reporter: Wei Zhu
>            Assignee: Marcus Eriksson
>             Fix For: 1.2.7
>
>         Attachments: 0001-CASSANDRA-5342-wip.patch, 
> 0001-CASSANDRA-5342-wip-v2.patch, Screen Shot 2013-03-13 at 12.05.08 PM.png
>
>
> We are using LCS and have total of 38000 SSTables for one CF. During LCS, 
> there could be over a thousand SSTable involved. All those SSTable IDs are 
> stored in ancestors field of SSTableMetatdata for the new table. In our case, 
> it consumes more than 1G of heap memory for those field. Put it in 
> perspective, the ancestors consume 2 - 3 times more memory than bloomfilter 
> (fp = 0.1 by default) in LCS. 
> We should remove those ancestors from SSTableMetadata after the compaction is 
> finished and the old SSTable is removed. It  might be a big deal for Sized 
> Compaction since there are small number of SSTable involved. But it consumes 
> a lot of memory for LCS. 
> At least, we shouldn't load those ancestors to the memory during startup if 
> the files are removed. 
> I would love to contribute and provide patch. Please let me know how to 
> start. 

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to