[
https://issues.apache.org/jira/browse/YUNIKORN-572?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17301397#comment-17301397
]
Wilfred Spiegelenburg commented on YUNIKORN-572:
------------------------------------------------
This is partially related to YUNIKORN-567.
Investigations for that Jira shows there is slightly more that needs to be done
but to clear things up: we should not hold the partition write lock while we
are processing application(s). Read locks are fine as multiple read locks can
be taken and things progress but scheduling takes a write lock on the
application and needs a read lock to list the nodes on a partition. So when
holding a partition write lock any scheduling action on an application could
cause the simple deadlock:
* other partition action: partition WRITE -> application READ
* scheduling action application WRITE -> partition READ
Partition write locks are limited to a subset of actions directly manipulating
the partition data.
Adding comments to the partition to clarify things.
> terminated app move causes deadlock
> -----------------------------------
>
> Key: YUNIKORN-572
> URL: https://issues.apache.org/jira/browse/YUNIKORN-572
> Project: Apache YuniKorn
> Issue Type: Bug
> Components: core - scheduler
> Reporter: Wilfred Spiegelenburg
> Assignee: Wilfred Spiegelenburg
> Priority: Major
>
> PR #250 for the placeholder cleanup introduced a possible dead lock.
> When moving a terminated app from the queue the queue is unlinked from the
> app. Before that happens we make sure that the queue tracking is up to date.
> This requires an application lock.
> The move used to take a lock on the partition and did all its work in one go.
> With the change to remove the queue from the app we now lock the app inside
> the partition lock. Since the app is part of the active list until the move
> is done scheduling might check the app too. This could lead to a dead lock.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
