hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jeffrey Zhong (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-9736) Alow more than one log splitter per RS
Date Thu, 14 Nov 2013 23:17:20 GMT

    [ https://issues.apache.org/jira/browse/HBASE-9736?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13823094#comment-13823094
] 

Jeffrey Zhong commented on HBASE-9736:
--------------------------------------

My thinking is to allow a user to configure a hard max split tasks per RS(N0). While each
RS internally calculates a run-time optimal number N1 = num of outstanding split log tasks
/ num of available RSs. SplitLogWorker use MIN( N1, N0) to spawn concurrent split log tasks.
In addition, upon a successful task grap, SplitLogWorker will sleep a little bit(yield) to
give other RSs a chance to take rest tasks in order to evenly distribute the work. 

> Alow more than one log splitter per RS
> --------------------------------------
>
>                 Key: HBASE-9736
>                 URL: https://issues.apache.org/jira/browse/HBASE-9736
>             Project: HBase
>          Issue Type: Improvement
>          Components: MTTR
>            Reporter: stack
>            Priority: Critical
>
> IIRC, this is an idea that came from the lads at Xiaomi.
> I have a small cluster of 6 RSs and one went down.  It had a few WALs.  I see this in
logs:
> 2013-10-09 05:47:27,890 DEBUG org.apache.hadoop.hbase.master.SplitLogManager: total tasks
= 25 unassigned = 21
> WAL splitting is held up for want of slots out on the cluster to split WALs.
> We need to be careful we don't overwhelm the foreground regionservers but more splitters
should help get all back online faster.



--
This message was sent by Atlassian JIRA
(v6.1#6144)

Mime
View raw message