[
https://issues.apache.org/jira/browse/HADOOP-2587?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12558515#action_12558515
]
Billy Pearson commented on HADOOP-2587:
---------------------------------------
The latest patch works good.
>From what I am seeing now is once a split is needed it waits until the next
>memcache flush is done and checks if compaction is running on that region if
>so then it waits for the next memcache flush to see if the compaction is done
>and so on and so on.
The only problem I see with this is
On a one region taking many min's to finsh compact and under heavy updates the
next compaction will start right after the first one is done unless you are
completely luck to finsh a compaction and a memcache flush at the same time
before the next compaction starts then it will just keep delayed the split.
Which is fine with me at this point the only time something like that would
happen is if there is only one region on a region server getting updated
heavily and its needing to compact over and over again with out the updates
slowing.
So right now the only improvement I can thank of is if the spliter can block a
new compaction from starting while waiting for the next memcache flush to
happen to get that lock. But if you can not do that then is no problem right
now the split will work as is and will split once a heave update load on a
region has slowed and the compaction has finshed.
Again this bug only happens when the region server has no other region to
compact and the region it has is under heave update load.
> Splits getting blocked by compactions causeing region to be offline for the
> length of the compaction 10-15 mins
> ---------------------------------------------------------------------------------------------------------------
>
> Key: HADOOP-2587
> URL: https://issues.apache.org/jira/browse/HADOOP-2587
> Project: Hadoop
> Issue Type: Bug
> Components: contrib/hbase
> Affects Versions: 0.16.0
> Environment: hadoop subversion 611087
> Reporter: Billy Pearson
> Assignee: Jim Kellerman
> Fix For: 0.16.0
>
> Attachments: hbase-root-regionserver-PE1750-3.log, log.log,
> patch.txt, patch.txt
>
>
> The below is cut out of one of my region servers logs full log attached
> What is happening is there is one region on a this region server and its is
> under heave insert load so compaction are back to back one one finishes a new
> one starts the problem starts when its time to split the region.
> A compaction starts just millsecs before the split starts blocking the split
> but the split closes the region before the compaction is finished. Causing
> the region to be offline until the compaction is done. Once the compaction is
> done the split finishes and all is returned to normal but this is a big
> problem for production if the region is offline for 10-15 mins.
> The solution would be not to let the split thread to issue the below line
> while a compaction on that region is happening.
> 2008-01-11 16:22:01,247 DEBUG org.apache.hadoop.hbase.HRegionServer:
> webdata,,1200085987488 closing (Adding to retiringRegions)
> The only time I have seen this bug is when there is only one region on a
> region server because if more then one then the compaction happens to the
> other region(s) after the first one is done compaction and the split can do
> what it needs on the first region with out getting blocked.
> {code}
> 2008-01-11 16:22:01,020 INFO org.apache.hadoop.hbase.HRegion: compaction
> completed on region webdata,,1200085987488. Took 16mins, 10sec
> 2008-01-11 16:22:01,020 DEBUG org.apache.hadoop.hbase.HStore: compaction for
> HStore webdata,,1200085987488/size needed.
> 2008-01-11 16:22:01,020 DEBUG org.apache.hadoop.hbase.HRegion:
> 1773667150/size needs compaction
> 2008-01-11 16:22:01,021 INFO org.apache.hadoop.hbase.HRegion: starting
> compaction on region webdata,,1200085987488
> 2008-01-11 16:22:01,021 DEBUG org.apache.hadoop.hbase.HStore: started
> compaction of 14 files using
> /gfs_storage/hadoop-root/hbase/hregion_1773667150/compaction.dir/hregion_1773667150/size
> for webdata,,1200085987488/size
> 2008-01-11 16:22:01,123 DEBUG org.apache.hadoop.hbase.HRegion: Started
> memcache flush for region webdata,,1200085987488. Size 31.2m
> 2008-01-11 16:22:01,232 INFO org.apache.hadoop.hbase.HRegion: Splitting
> webdata,,1200085987488 because largest aggregate size is 100.7m and desired
> size is 64.0m
> 2008-01-11 16:22:01,247 DEBUG org.apache.hadoop.hbase.HRegionServer:
> webdata,,1200085987488 closing (Adding to retiringRegions)
> ...
> lots of NotServingRegionException's
> ...
> 2008-01-11 16:32:59,876 INFO org.apache.hadoop.hbase.HRegion: compaction
> completed on region webdata,,1200085987488. Took 10mins, 58sec
> ...
> 2008-01-11 16:33:02,193 DEBUG org.apache.hadoop.hbase.HRegion: Cleaned up
> /gfs_storage/hadoop-root/hbase/hregion_1773667150/splits true
> 2008-01-11 16:33:02,194 INFO org.apache.hadoop.hbase.HRegion: Region split of
> webdata,,1200085987488 complete; new regions: webdata,,1200090121237,
> webdata,com.tom.ent/2008-01-04/0PGM/09034104.html:http,1200090121239. Split
> took 11mins, 0sec
> 2008-01-11 16:33:02,227 DEBUG
> org.apache.hadoop.hbase.HConnectionManager$TableServers: No servers for
> .META.. Doing a find...
> 2008-01-11 16:33:02,283 DEBUG
> org.apache.hadoop.hbase.HConnectionManager$TableServers: Found 1 region(s)
> for .META. at address: 10.0.0.4:60020, regioninfo: regionname: -ROOT-,,0,
> startKey: <>, encodedName(70236052) tableDesc: {name: -ROOT-, families:
> {info:={name: info, max versions: 1, compression: NONE, in memory: false, max
> length: 2147483647, bloom filter: none}}}
> 2008-01-11 16:33:02,284 INFO org.apache.hadoop.hbase.HRegionServer: Updating
> .META. with region split info
> 2008-01-11 16:33:02,290 DEBUG org.apache.hadoop.hbase.HRegionServer:
> Reporting region split to master
> 2008-01-11 16:33:02,291 INFO org.apache.hadoop.hbase.HRegionServer: region
> split, META update, and report to master all successful. Old
> region=webdata,,1200085987488, new regions: webdata,,1200090121237,
> webdata,com.tom.ent/2008-01-04/0PGM/09034104.html:http,1200090121239
> {code}
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
