[
https://issues.apache.org/jira/browse/HBASE-722?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jim Kellerman updated HBASE-722:
--------------------------------
Fix Version/s: (was: 0.18.0)
0.19.0
> Shutdown and Compactions
> ------------------------
>
> Key: HBASE-722
> URL: https://issues.apache.org/jira/browse/HBASE-722
> Project: Hadoop HBase
> Issue Type: Improvement
> Components: regionserver
> Reporter: Billy Pearson
> Priority: Trivial
> Fix For: 0.19.0
>
>
> Currently I thank the region server closes the region spits in a set order
> this same order is used in the compaction/spit thread after a startup and all
> regions need compaction.
> This causes the shutdown time to take a vary long time depending on what
> region the server is compacting.
> I have seen my cluster take more then 60 mins to shutdown waiting for
> several regions to finish compaction.
> The problem I am seeing is say the compaction thread is working on a on
> region 2 of 4 in the list.
> When the region server get the MSG_REGIONSERVER_QUIESCE from the master it
> closes in order from 1-4
> So it closes region 1 and waits for region 2 to finish compaction before
> closing it.
> The problem is it never closes region split 3 or 4 while waiting for region 2
> to complete the compaction.
> So example say 3 and 4 regions are waiting to be compacted also. When region
> 2 is done region 3 start compaction within milliseconds of 2's finished
> compaction.
> This happens faster then the region server can close region 3 so the region
> server hangs around compacting regions until it run out of open regions
> needing compact.
> With a lot of regions this could hang some region server a long time to
> Shutdown the cluster.
> Solutions 1 or 2 below will work I thank
> 1. Close all regions not currently getting compaction so when the compaction
> thread complete it has no other regions open to compact
> 2. Empty the compaction que so it has no other regions to compact when done
> with the current one.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
