Meng Zhu created MESOS-8850:
-------------------------------
Summary: Race between master and allocator when destroying shared
volume could lead to sorter check failure.
Key: MESOS-8850
URL: https://issues.apache.org/jira/browse/MESOS-8850
Project: Mesos
Issue Type: Bug
Components: allocation, master
Reporter: Meng Zhu
When destroying shared volume, master first rescinds offers that contain the
shared volume and then apply the destroy operation. This process involves
interaction between the master and allocator actor. The following race could
arise:
1. Framework1 and framework2 are each offered a shared disk;
2. Framework2 asks the master to destroy the shared disk;
3. Master rescinds framework1's offer that contains the shared disk;
4. `allocator->recoverResources` is called to recover framework1’s offered
resources in the allocator;
5. [Race] Allocator shortly allocates resources to framework1. The allocation
contains the shared disk that just got recovered which has not been destroyed
at the moment. Allocator invokes `offerCallback` which dispatches to the master;
6. Master continues the destroy operation and calls
`allocator->updateAllocation` to notify the allocator to transform the shared
disk to regular reserved disk;
7. Master processes the `offerCallback` dispatched in step5 and offered the
shared disk to framework1.
At this point, the same disk resource appears in two different places: one
shared offered to framework1, one not shared currently hold by framework2 (soon
to be recovered).
One aftermath is that:
Framework2’s resources get recovered which includes the (now regular reserved)
disk resource.
Later, when recovering framework1’s resources which contains the shared disk,
the sorter finds that allocated resources on the agent do not contain that
shared disk (because in step 5 when offering the shared disk, the allocator did
not increase the total allocated resources as framework2 was also holding the
shared disk. We only add shared resource to allocated only when it is allocated
the first time).
This will lead to check failure in sorter:
https://github.com/apache/mesos/blob/master/src/master/allocator/sorter/drf/sorter.cpp#L480
Moving offer management to the allocator could definitely eliminate this race.
Without that, we will need to add extra synchronizations.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
