[ 
https://issues.apache.org/jira/browse/SPARK-60862?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Jungtaek Lim updated SPARK-60862:
---------------------------------
    Affects Version/s: 4.4.0

> Watermark propagation through reused exchanges uses zero watermark
> ------------------------------------------------------------------
>
>                 Key: SPARK-60862
>                 URL: https://issues.apache.org/jira/browse/SPARK-60862
>             Project: Spark
>          Issue Type: Bug
>          Components: Structured Streaming
>    Affects Versions: 5.0.0, 4.4.0
>            Reporter: Siying Dong
>            Assignee: Siying Dong
>            Priority: Major
>              Labels: pull-request-available
>
> ReusedExchangeExec is a leaf physical plan node whose underlying exchange is 
> stored in its
> child field rather than in the normal children sequence. 
> PropagateWatermarkSimulator does not
> currently unwrap this node when traversing the plan or resolving an input 
> watermark.
> As a result, when equivalent streaming aggregation branches reuse an 
> exchange, the direct
> branch receives the correct event-time watermark while the reused branch 
> remains pinned at
> zero. Downstream stateful operators also receive zero. In append mode, the 
> affected branch may
> never emit closed windows and its state can grow without bound, causing wrong 
> results and
> unbounded state growth.
> The fix treats ReusedExchangeExec as a transparent wrapper in both watermark 
> traversal and
> input-watermark lookup, with a guard against re-traversing a hidden child 
> already visited
> through the canonical exchange. A streaming SQL regression test covers a 
> union of two windowed
> aggregation branches and verifies that both branches emit their closed 
> windows after the
> watermark advances.
> Pull request: https://github.com/apache/spark/pull/60073



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to