[
https://issues.apache.org/jira/browse/HBASE-15959?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jingcheng Du updated HBASE-15959:
---------------------------------
Comment: was deleted
(was: Thanks [~appy], you are right. The patch only fixes the issue in this
test but not fix the uncovered bug.
bq. I don't know much about MOB, but from the looks, it seems that the test has
uncovered a bug here
What I gave in the example is not only for mob, it is what the compactor
behaves in the master branch for all kinds of compactions.
Now in the master code, all (not only for mob) compactors won't archive the
old/compacted store files, instead a region server chore does it. But this
cleaning is not done when the region is close. I guess the WAL can help redo
this when the region starts, but in the test I didn't see this, The chaos is
there after the region is restarted ( I did this by disabling/enabling tables).
Fortunately the scan results will be right in both non-mob and mob tables, but
metrics in mob tables might be wrong because of the chaos. I am looking into
why the cleaning is not done before close.)
> Fix flaky test TestRegionServerMetrics.testMobMetrics
> -----------------------------------------------------
>
> Key: HBASE-15959
> URL: https://issues.apache.org/jira/browse/HBASE-15959
> Project: HBase
> Issue Type: Bug
> Reporter: Appy
> Assignee: huaxiang sun
> Attachments: HBASE-15959-v001.patch, HBASE-15959-v002.patch,
> HBASE-15959-v003.patch
>
>
> It flakes
> [here|https://github.com/apache/hbase/blob/b557f0bec62a48753e5d01d7a47f3c9e5a6b3ee8/hbase-server/src/test/java/org/apache/hadoop/hbase/regionserver/TestRegionServerMetrics.java#L460].
> There are two weird things i identified:
> 1. In second compaction,
> [scanner|https://github.com/apache/hbase/blob/b557f0bec62a48753e5d01d7a47f3c9e5a6b3ee8/hbase-server/src/main/java/org/apache/hadoop/hbase/mob/DefaultMobStoreCompactor.java#L173]
> has 10 storefiles. Shouldn't there be 6? 5 from recent flushes and 1 from
> earlier compaction. Probably because mob cleaner doesn't clean old hfiles.
> Does this needs fixing?
> 2. Across runs, same cell (ie. same key) may or may not be considered mob
> reference cell.
> [here|https://github.com/apache/hbase/blob/b557f0bec62a48753e5d01d7a47f3c9e5a6b3ee8/hbase-server/src/main/java/org/apache/hadoop/hbase/mob/DefaultMobStoreCompactor.java#L213].
> This at least happens with row keys 0 - 4 (which got compacted earlier).
> [~jmhsieh] Any ideas why this would happen.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
