hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Dave Latham (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-2770) Major compactions from shell may not major compact all families
Date Tue, 13 Jul 2010 19:47:56 GMT

    [ https://issues.apache.org/jira/browse/HBASE-2770?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12887962#action_12887962

Dave Latham commented on HBASE-2770:

It was definitely a headache for us during our data center migration, but I've got no immediate
need for it if you're wanting to get 0.20.6 out and this is blocking the way.

You can put a cluster into read-only mode and make sure that hbase doesn't touch the store
files while copying them away by setting hbase.hregion.majorcompaction to a very high value
to workaround this issue.

> Major compactions from shell may not major compact all families
> ---------------------------------------------------------------
>                 Key: HBASE-2770
>                 URL: https://issues.apache.org/jira/browse/HBASE-2770
>             Project: HBase
>          Issue Type: Bug
>    Affects Versions: 0.20.4
>            Reporter: Dave Latham
>            Priority: Critical
>             Fix For: 0.20.6, 0.90.0
> As part of a data center migration, I initiated a major_compaction request on all tables
from the shell.  A few hours later, all the region servers in the cluster appeared to have
completed the compactions and all compactionQueue metrics were back to 0.  However, some column
families of some regions had not actually done a major compaction.
> Digging through logs and code, it looks like the following happened.  The shell makes
a major compaction request which sets HRegion.forceMajorCompaction to true for every region.
 Periodically, the HRegionServer.MajorCompactionChecker checks to see if a major compaction
is needed in any family's store.  If so, calls CompactSplitThread.compactionRequested which
ends up setting the region forceMajorCompaction to false, even if it is already in the compaction
queue and set to true.  Then, when that region comes off the queue to be compacted, each family/store
separately checks for whether it should do a major compaction, so some families may not do
> (This is not good if, for example, you're doing a DistCp of the hbase dir and later on
the cluster decides to do a compaction on those files and deletes ones the DistCp job is looking
for, causing it to fail.)

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message