[
https://issues.apache.org/jira/browse/HBASE-1597?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12727000#action_12727000
]
Jonathan Gray commented on HBASE-1597:
--------------------------------------
Good point, Andrew. Somehow I forgot about an issue I actually filed,
HBASE-1521.
Really what I didn't like in my patch was how I re-instantiate the Reader as a
CompactionReader, since I can't cast up to it. I did it in a way that just
assigns all the internal variables so there's not much real overhead besides
the instantiation. Seems like there could be a prettier way but it works and
performance impact should be negligible.
Thanks.
> Prevent unnecessary caching of blocks during compactions
> --------------------------------------------------------
>
> Key: HBASE-1597
> URL: https://issues.apache.org/jira/browse/HBASE-1597
> Project: Hadoop HBase
> Issue Type: Bug
> Affects Versions: 0.20.0
> Reporter: Jonathan Gray
> Assignee: Jonathan Gray
> Fix For: 0.20.0
>
> Attachments: HBASE-1597-v1.patch
>
>
> When running any kind of compaction, we read every block of every storefile
> being compacted into the block cache.
> We would like to reuse any already cached blocks, if available, but otherwise
> we should not bog down the LRU with unnecessary blocks.
> This is not as bad as it was with the old LRU because the latest LRU
> implementation (HBASE-1460) is scan-resistant. This ensures that we are not
> causing massive eviction of the blocks that are being read multiple times or
> from in-memory tables. However, this does add to the GC-woes of an import
> because each block gets further referenced, and for longer periods of time.
> There is also overhead in running the LRU evictions.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
