[
https://issues.apache.org/jira/browse/YARN-7899?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Botong Huang updated YARN-7899:
-------------------------------
Attachment: YARN-7899.v2.patch
> [AMRMProxy] Stateful FederationInterceptor for pending requests
> ---------------------------------------------------------------
>
> Key: YARN-7899
> URL: https://issues.apache.org/jira/browse/YARN-7899
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Botong Huang
> Assignee: Botong Huang
> Priority: Major
> Attachments: YARN-7899.v1.patch, YARN-7899.v2.patch
>
>
> Today FederationInterceptor (in AMRMProxy for YARN Federation) is stateless
> in terms of pending (outstanding) requests. Whenever AM issues new requests,
> FI simply splits and sends them to sub-cluster YarnRMs and forget about them.
> This JIRA attempts to make FI stateful so that it remembers the pending
> requests in all relevant sub-clusters. This has two major benefits:
> 1. It is a prerequisite for FI to be able to cancel pending request in one
> sub-cluster and re-send it to other sub-clusters. This is needed for load
> balancing and to fully comply with the relax locality fallback to ANY
> semantic. When we send a request to one sub-cluster, we have effectively
> restrained the allocation for this request to be within this sub-cluster
> rather than everywhere. If the cluster capacity in this sub-cluster for this
> app is full or this YarnRM is overloaded and slow, the request will be stuck
> there for a long time even if there is free capacity in other sub-clusters.
> We need FI to remember and adjust the pending requests on the fly.
> 2. This makes pending request recovery easier when YarnRM fails over. Today
> whenever one sub-cluster RM fails over, in order to recover lost pending
> requests for this sub-cluster,
> we have to propagate the ApplicationMasterNotRegisteredException from the
> YarnRM back to AM, triggering a full pending resend from AM. This contains
> pending for not only the failing-over sub-cluster, but everyone. Since our
> split-merge (AMRMProxyPolicy) does not guarantee idempotency, the same
> request we sent to sub-cluster-1 earlier might be resent to sub-cluster-2. If
> both these YarnRMs have not failed over, they will both allocate for this
> request, leading to over-allocation. Also, these full pending resends also
> puts unnecessary load on every YarnRM in the cluster everytime one YarnRM
> fails over. With stateful FederationInterceptor, since we remember pending
> requests we have sent out earlier, we can shield the
> ApplicationMasterNotRegisteredException for AM and resend the pending only to
> the failed over YarnRM. This eliminates over-allocation and minimizes the
> recovery overhead.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
