[jira] [Updated] (YARN-11992) Multi-node allocation discards an earlier reservation after later nodes are skipped

["ASF GitHub Bot (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/YARN-11992?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

ASF GitHub Bot updated YARN-11992:
----------------------------------
    Labels: pull-request-available  (was: )

> Multi-node allocation discards an earlier reservation after later nodes are skipped
> -----------------------------------------------------------------------------------
>
>                 Key: YARN-11992
>                 URL: https://issues.apache.org/jira/browse/YARN-11992
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: capacity scheduler, multi-node-placement
>            Reporter: Yuan Luo
>            Priority: Major
>              Labels: pull-request-available
>
> h3. Problem
> When multi-node placement is enabled, {{RegularContainerAllocator.allocate()}} continues to inspect other candidate nodes after one node returns {{RESERVED}}. If a later candidate is skipped, that later result overwrites the earlier reservation. The method then returns the skip result and the valid reservation is lost.
> The reservation must also be completed against the node that produced it. Using the last visited candidate would associate the reserved container with the wrong node.
> h3. Reproduction
> # Use a candidate set containing at least two nodes.
> # The first node can fit the request in total capacity but lacks currently available resources, so allocation returns {{RESERVED}}.
> # A later node cannot satisfy the same request and returns a skip result.
> Actual result: the allocator returns the later skip result and does not create the earlier reservation.
> Expected result: after candidate traversal is exhausted, the allocator should create the most recent valid reservation on the node that produced it. If a later allocation attempt reaches {{doAllocation()}}, its result must be returned directly instead of falling back to a stale reservation.
> h3. Proposed fix
> Track both the most recent reservation result and its corresponding node. If traversal finishes without an allocation, complete that reservation on the saved node. Return the result of an allocation attempt directly so a failed allocation cannot incorrectly restore an older reservation.
> h3. Tests
> Add focused coverage for:
> * reservation followed by a skipped candidate;
> * multiple reservations, where the latest reservation and its node win;
> * a failed allocation after an earlier reservation, which must not fall back to the stale reservation.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org