[
https://issues.apache.org/jira/browse/HBASE-11397?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14040142#comment-14040142
]
Ted Yu commented on HBASE-11397:
--------------------------------
Thanks for trying out stripe compaction.
In the future, please attach server log in text format.
> When merging expired stripes, we need to create an empty file to preserve
> metadata.
> -----------------------------------------------------------------------------------
>
> Key: HBASE-11397
> URL: https://issues.apache.org/jira/browse/HBASE-11397
> Project: HBase
> Issue Type: Bug
> Components: Compaction
> Affects Versions: 0.98.2
> Environment: jdk1.7.0_45, hadoop-cdh5, hbase-0.98.2
> Reporter: Victor Xu
> Attachments: HBASE-11397-AssertionError.png, HBASE-11397-HDFS.png,
> HBASE-11397-RS-Log.png, HBASE-11397-Stripe-Info.png, HBASE-11397.patch
>
>
> Stripe Compaction is a good feature in 0.96 and 0.98. But when I used it in a
> heavy-write non-uniform row keys scenario(e.g. time dimension in a key), I
> came across some problems.
> I made my stripes split at the size of 2G(hbase.store.stripe.sizeToSplit=2G),
> and soon there were tens of them. It was true that only the last stripe
> receiving the new keys kept compacting - old data didn't compact as much, or
> at all. However, the old stripes were still there when they all expired. I
> checked the source code and found that when compacting expired stripes, the
> StoreScanner may return no KVs so that SizeMultiWriter.append() is never
> called. That's to say, NO NEW FILE WILL BE CREATED.
> My solution is to create an empty file to preserve metadata at the end of the
> SizeMultiWriter.commitWritersInternal().
--
This message was sent by Atlassian JIRA
(v6.2#6252)
