[jira] Reopened: (HBASE-921) region close and open processed out of order; makes for disagreement between master and regionserver on region state

["Andrew Purtell (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HBASE-921?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Andrew Purtell reopened HBASE-921:
----------------------------------

      Assignee:     (was: Jim Kellerman)

I believe I saw an instance of this issue again today. 

What about Jim Firby's original suggestion: "Clients should be able to say "I've asked for a regions location 10 times now and Mr. Master, you've given me the same response ten times in a row and each time, the answer was wrong. Revisit any notion that said region is at said location". Mr. Master would then go off and do something drastic like close and reassign the region."

Or the Master can sanity check on its own. If my latest patch to HBASE-1018 goes in, the master can look at the HServerLoad and note that not all expected regions are found there. 

> region close and open processed out of order; makes for disagreement between master and regionserver on region state
> --------------------------------------------------------------------------------------------------------------------
>
>                 Key: HBASE-921
>                 URL: https://issues.apache.org/jira/browse/HBASE-921
>             Project: Hadoop HBase
>          Issue Type: Sub-task
>    Affects Versions: 0.18.0
>            Reporter: stack
>            Priority: Blocker
>             Fix For: 0.18.1, 0.19.0
>
>         Attachments: 921-0.18.0.patch
>
>
> Master assigns region X successfully.  It then decides to close it because it wants it opened elsewhere as part of region rebalancing.  Both the open and close operations are reported back to the master.  Both have operation processing components that are added to the todo list to be processed in another thread outside of the master's main loop.
> The close operation does the bulk of its work inline with the master main processing loop.  Its todo component does some work if the region is offlined but otherwise nothing of consequence whereas the open in its todo does the important meta catalog table update with the new location information.
> Its been fairly common here on our cluster where the master todo queue is occupied processing the shutdown of a regionserver.  It takes a long time to process the shutdown of a regionserver when thousands of regions   This latter delays the processing of the open and close todos.  In effect the open is running after the close.  The region goes into limbo.  Only a restart of the 'hosting' regionserver 'fixes' this state.
> This is a particular case of the general HBASE-543 issue.  Its happening alot here on our cluster so will hack up a fix for this and get it into TRUNK and backport it to 0.18.1.
> Jim Firby here had a good idea for conditions like this.  Clients should be able to say "I've asked for a regions location 10 times now and Mr. Master, you've given me the same response ten times in a row and each time, the answer was wrong.  Revisit any notion that said region is at said location".  Mr. Master would then go off and do something drastic like close and reassign the region.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.