[jira] [Commented] (HBASE-21806) add an option to roll WAL on very slow syncs

["Duo Zhang (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-21806?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16755559#comment-16755559 ] 

Duo Zhang commented on HBASE-21806:
-----------------------------------

One of my goal for introducing AsyncFSWAL is to set a smaller timeout when writing WAL, so if there are slow DNs we will fail to sync sonn and roll a new WAL...

But anyway, the proposal here may still be useful, as for example, we can set the timeout to 5 seconds, and set a 2 seconds threshold, if we haven't got response after 5 seconds, the sync will fail, if the sync can come back but still spends more than 2 seconds, we roll a new WAL.

> add an option to roll WAL on very slow syncs
> --------------------------------------------
>
>                 Key: HBASE-21806
>                 URL: https://issues.apache.org/jira/browse/HBASE-21806
>             Project: HBase
>          Issue Type: Bug
>            Reporter: Sergey Shelukhin
>            Assignee: Sergey Shelukhin
>            Priority: Major
>         Attachments: HBASE-21806.patch
>
>
> In large heterogeneous clusters sometimes a slow datanode can cause WAL syncs to be very slow. In this case, before the bad datanode recovers, or is discovered and repaired, it would be helpful to roll WAL on a very slow sync to get a new pipeline.
> Otherwise the slow WAL will impact write latency for a long time (slow writes result in less writes result in the WAL not being rolled for longer)



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)