hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ted Yu (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-19056) TestCompactionInDeadRegionServer is top of the flakies charts!
Date Sun, 26 Nov 2017 23:39:00 GMT

    [ https://issues.apache.org/jira/browse/HBASE-19056?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16266255#comment-16266255
] 

Ted Yu commented on HBASE-19056:
--------------------------------

I spent a bit time on this test and the following is what I found.

For current test, most likely the assertion fails for the second parameterized run where table
region and hbase:meta were on same server before the watcher.getRecoverableZooKeeper().delete()
call.

I also verified the un-writable store condition which causes the compaction request to come
back without actually triggering the compaction, leading to fail() assertion.

In the modified tests, I would detect whether table region is on the same server as hbase:meta.
If so, I would move the region to the other server before issuing the watcher.getRecoverableZooKeeper().delete()
call.

I have looped both tests 9 times locally which passed.

>  TestCompactionInDeadRegionServer is top of the flakies charts!
> ---------------------------------------------------------------
>
>                 Key: HBASE-19056
>                 URL: https://issues.apache.org/jira/browse/HBASE-19056
>             Project: HBase
>          Issue Type: Bug
>          Components: test
>            Reporter: stack
>         Attachments: 19056.v5.txt
>
>
> The test came in recently as part of HBASE-17712 "Remove/Simplify the logic of RegionScannerImpl.handleFileNotFound"
> [~Apache9] when you have a chance, help me out. I was going to just remove the test since
it made no sense to me but then I saw you wrote it  (smile).
> When the region.compact(true); is called on the end, what is supposed to be going on?
> When I trace, the compact is not done because the Region is not writeEnabled (we check
if Region is writeEnabled down in Store before we go ahead and compact). So, I thought the
problem was that the region reference was stale because it came from the rsToSuspend which
had just been killed.
> After a while, I figured that you intend the region reference to be stale so you can
try an append AFTER the WAL has been taken over by WAL splitter.
> But the writeEnabled flag is set so compactions don't run. I tried unsetting this flag
and closed flags and but compaction won't run.
> Was this your intent sir? If so, I'll work w/ it np. Just looking for clarity. Thanks.




--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message