[
https://issues.apache.org/jira/browse/FLINK-15224?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16995723#comment-16995723
]
Zhu Zhu commented on FLINK-15224:
---------------------------------
The managed memory weight is not in ResourceProfile in 1.10 so it would not be
a problem.
And I just realized that this may not be a issue, as long as we can ensure a
shared slot is large enough to host all possible task instances (one instance
per JobVertex in the corresponding SlotSharingGroup). So maybe adding a sanity
check for it would be enough in 1.11 (if we will not change the slot sharing
mechanism then).
> Resource requirements are not respected when fulfilling a slot request with
> unresolvedRootSlots from a SlotSharingManager
> -------------------------------------------------------------------------------------------------------------------------
>
> Key: FLINK-15224
> URL: https://issues.apache.org/jira/browse/FLINK-15224
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Coordination
> Affects Versions: 1.11.0
> Reporter: Zhu Zhu
> Priority: Major
>
> In {{SchedulerImpl#allocateMultiTaskSlot}}, if a slot request cannot be
> fulfilled immediately with a resolved root slot(MultiTaskSlot that is
> fulfilled by an allocated slot) or with available slots, it will be assigned
> to a random unresolved root slot. It does not do resource requirements check
> in this case, so a large task slot can be assigned to a small shared slot
> (unresolved root slot) and when the shared slot received its physical slot
> offer, it will be recognized as oversubscribing and the slot would be
> released and related tasks would fail.
> It's not a problem for now since specified resources are not used yet, but
> can be a problem in the future when we are to support specified resources.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
