[
https://issues.apache.org/jira/browse/HBASE-18128?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16040085#comment-16040085
]
Jingyun Tian commented on HBASE-18128:
--------------------------------------
[~anoop.hbase] [~tedyu]
Just submitted a patch and unit test for this issue. Please have a look.
I found there is a method OutputSink.keepRegionEvent(Entry entry) check if we
should skip the meta cell on the master branch, so I wrote the patch base on
the method. Should I back port it to earlier releases?
The unit test are almost same with
TestWALSplit.testSplitLeavesCompactionEventsEdits(), the only difference is the
seqNum of compaction marker I set is -2. Since the default
lastFlushedSequenceId is -1, it will skip the compaction marker if we don't
apply the patch.
> compaction marker could be skipped
> -----------------------------------
>
> Key: HBASE-18128
> URL: https://issues.apache.org/jira/browse/HBASE-18128
> Project: HBase
> Issue Type: Improvement
> Components: Compaction, regionserver
> Reporter: Jingyun Tian
> Assignee: Jingyun Tian
> Attachments: HBASE-18128-master.patch, TestCompactionMarker.java
>
>
> The sequence for a compaction are as follows:
> 1. Compaction writes new files under region/.tmp directory (compaction output)
> 2. Compaction atomically moves the temporary file under region directory
> 3. Compaction appends a WAL edit containing the compaction input and output
> files. Forces sync on WAL.
> 4. Compaction deletes the input files from the region directory.
> But if a flush happened between 3 and 4, then the regionserver crushed. The
> compaction marker will be skipped when splitting log because the sequence id
> of compaction marker is smaller than lastFlushedSequenceId.
> {code}
> if (lastFlushedSequenceId >= entry.getKey().getLogSeqNum()) {
> editsSkipped++;
> continue;
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
