[
https://issues.apache.org/jira/browse/HBASE-30385?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18115958#comment-18115958
]
huginn commented on HBASE-30385:
--------------------------------
PR: https://github.com/apache/hbase/pull/8658
> Avoid scheduling duplicate system region moves
> ----------------------------------------------
>
> Key: HBASE-30385
> URL: https://issues.apache.org/jira/browse/HBASE-30385
> Project: HBase
> Issue Type: Bug
> Components: master
> Affects Versions: 2.4.11
> Reporter: huginn
> Priority: Major
> Labels: pull-request-available
>
> When a newer-version RegionServer joins a mixed-version cluster,
> AssignmentManager.checkIfShouldMoveSystemRegionAsync() can schedule the same
> non-meta system-region move more than once when multiple excluded servers are
> processed.
> The method accumulates RegionPlans in a list while iterating excluded
> servers, but currently iterates over the accumulated list inside that same
> loop. A plan collected from an earlier server is therefore submitted again
> when the next excluded server is processed. The second submission can hit the
> in-transition guard and emit an exception such as:
> {code}
> ... is currently in transition; pid=...
> {code}
> This can also prevent the remaining plans from being submitted.
> The fix is to submit the collected plans after the server loop, and use
> plan.getSource() in the log so the source server matches the plan.
> Reproduction and validation:
> * Use a mixed-version cluster with at least two excluded servers containing
> system regions.
> * Source review against Apache master identifies the duplicated iteration.
> * No unit test is included; this is a small control-flow and logging
> correction.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
