[
https://issues.apache.org/jira/browse/HBASE-8081?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Devaraj Das updated HBASE-8081:
-------------------------------
Attachment: 7213-0.94-2.patch
Good news on both fronts -
1. This won't create upgrade/downgrade issues. The new code creates the log
files with .meta extensions for the meta regions. The new code can handle logs
that have meta and other regions' edits mingled together. The old code lists
all the files in the logs directory and tries to split them (so it would also
split the files with .meta extensions).
2. The performance is good. I ran the LoadTestTool on a cluster with two RSs.
After the test runs for a while (around 30 or so minutes), I would kill the
server hosting the .META. Without the patch, I noticed the master reports takes
around ~20-30 seconds to do the log splitting ("finished splitting" log line).
With the patch, it first tries to do the splitting of the meta logs and that is
complete in ~1-1.5 seconds. Meta tables come back online that much sooner.
[FYI: the cmd I ran for testing: bin/hbase
org.apache.hadoop.hbase.util.LoadTestTool -write 10:100:20 -read 10:1 -num_keys
1000000]
The attached patch is what I tested. I have fixed a bug in the process (race
condition to do with hlog creation if both _ROOT_ and .META. tables are in the
same RS). Put in a comment to that effect in HRegionServer.getMetaWAL. I think
I have good confidence in the patch (i ran a lot of loadtestool runs with the
patch).
> Backport HBASE-7213 (separate hlog for meta tables) to 0.94
> -----------------------------------------------------------
>
> Key: HBASE-8081
> URL: https://issues.apache.org/jira/browse/HBASE-8081
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.94.5
> Reporter: Devaraj Das
> Assignee: Devaraj Das
> Fix For: 0.94.7
>
> Attachments: 7213-0.94-2.patch, 7213-0.94.patch
>
>
> I am interested in backporting HBASE-7213 to 0.94. Helps to address more of
> the MTTR story. Offline discussion with Lars indicated he is interested as
> well.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
