Status: Patch Available
Previously we have encountered a serious problem in ResourceManager, we found that pending container number of one RM queue became negative after RM failed over. Since queues in RM are managed in hierarchical structure, the root queue's pending containers became negative at last, thus the scheduling process of the whole cluster became affected.
The version of both our RM server and AMRM client in our application are based on yarn 3.1, and we uses AMRMClientAsync#addSchedulingRequests() method in our application to request resources from RM.
After investigation, we found that the direct cause was numAllocations of some AMs' requests became negative after RM failed over. And there are at lease three necessary conditions:
(1) Use schedulingRequests in AMRM client, and the application set zero to the numAllocations for a schedulingRequest. In our batch job scenario, the numAllocations of a schedulingRequest could turn to zero because theoretically we can run a full batch job using only one container.
(2) RM failovers.
(3) Before AM reregisters itself to RM after RM restarts, RM has already recovered some of the application's containers assigned before.
Here are some more details about the implementation:
(1) After RM recovers, RM will send all alive containers to AM once it re-register itself through RegisterApplicationMasterResponse#getContainersFromPreviousAttempts.
(2) During registerApplicationMaster, AMRMClientImpl will removeFromOutstandingSchedulingRequests once AM gets ContainersFromPreviousAttempts without checking whether these containers have been assigned before. As a consequence, its outstanding requests might be decreased unexpectedly even if it may not become negative.
(3) There is no sanity check in RM to validate requests from AMs.
For better illustrating this case, I've written a test case based on the latest hadoop trunk, posted in the attachment. You may try case testAMRMClientWithNegativePendingRequestsOnRMRestart and testAMRMClientOnUnexpectedlyDecreasedPendingRequestsOnRMRestart .
To solve this issue, I propose to filter allocated containers before removeFromOutstandingSchedulingRequests in AMRMClientImpl during registerApplicationMaster, and some sanity checks are also needed to prevent things from getting worse.
More comments and suggestions are welcomed.