hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jean-Daniel Cryans (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-8171) Perf issue with multiple client threads trying to prefetch Region Cache from META
Date Thu, 21 Mar 2013 20:25:16 GMT

    [ https://issues.apache.org/jira/browse/HBASE-8171?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13609423#comment-13609423

Jean-Daniel Cryans commented on HBASE-8171:

We had a longish discussion about this too here: https://issues.apache.org/jira/browse/HBASE-6841

My opinion is that pre-fetching is bad for most use cases the way it is implemented and I
prefer to run without it in production.
> Perf issue with multiple client threads trying to prefetch Region Cache from META
> ---------------------------------------------------------------------------------
>                 Key: HBASE-8171
>                 URL: https://issues.apache.org/jira/browse/HBASE-8171
>             Project: HBase
>          Issue Type: Bug
>    Affects Versions: 0.89-fb, 0.94.3, 0.94.5
>            Reporter: Amitanand Aiyer
>            Assignee: Amitanand Aiyer
> HBaseClient maintains a cache of Regions to region locations that gets cleared whenever
the RS that the client talks to is unresponsive.
> This will result in the Client having to talk to META to fetch the region location before
doing the get/put.
> https://issues.apache.org/jira/browse/HBASE-2468 introduces the feature to prewarm the
cache, where the client gets some rows from META to warm the cache.
> This is a costly operation. And,  it looks like all client threads that block on the
regionLockObject, will go in and do the cache warmup; although the cache might have been already
warmed up by an earlier thread. 
> When there is a network issue going on, we see lots of Client threads blocking in locateRegionInMeta.
> The fix would be to check the cache again; before we do the prefetch after grabbing the
lock. Currently, the client threads look in the cache at the begining of the method. But a
lot could have changed while waiting for the regionLockObject.

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

View raw message