[
https://issues.apache.org/jira/browse/HBASE-5344?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13208089#comment-13208089
]
Phabricator commented on HBASE-5344:
------------------------------------
mbautin has commented on the revision "[jira] [HBASE-5344] [89-fb] Scan
unassigned region directory on master failover".
The bigger context of our current/planned changes in 89-fb master is as
follows. In 89-fb, region assignments happen as responses to RS -> master RPC,
and RSs communicate success of region open operations back to the master
through ZK. The master then writes the new assignments to META. ZK is the only
piece in the picture that could be considered a trusted highly-available source
of truth for the region assignment, if only it had all assignments. Currently
the region assignment can be obtained from the combination of META and ZK's
UNASSIGNED directory. We have a plan to move towards always having the full
assignment in ZK (the UNASSIGNED directory will change its meaning then) to
help guarantee that we never have a duplicate assignment and to have only one
source of truth for assignment. We will also keep writing the region assignment
to META for client backward-compatibility. Even though the master failover fix
does not depend on those planned changes, I thought it would b
e useful to mention them here.
Contacting all regionservers directly to get the region assignment is
probably useful as a sanity-check, but it is not scalable, and is subject to
unpredictable timeouts in the worst case. We would like to rely on ZK and (for
now) on META instead to recover the region assignment on master
startup/failure. Also, by the way, we are planning to unify master startup on a
fresh cluster start and failover and everything in between, and use the same
logic to build a coherent picture of region assignment.
REVISION DETAIL
https://reviews.facebook.net/D1605
> [89-fb] Scan unassigned region directory on master failover
> -----------------------------------------------------------
>
> Key: HBASE-5344
> URL: https://issues.apache.org/jira/browse/HBASE-5344
> Project: HBase
> Issue Type: Bug
> Reporter: Mikhail Bautin
> Assignee: Mikhail Bautin
> Attachments: D1605.1.patch
>
>
> In case the master dies after a regionserver writes region state as OPENED or
> CLOSED in ZK but before the update is received by master and written to meta,
> the new master that comes up has to pick up the region state from ZK and
> write it to meta. Otherwise we can get multiply-assigned regions.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
