[
https://issues.apache.org/jira/browse/BEAM-13541?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17466169#comment-17466169
]
Ahmet Altay commented on BEAM-13541:
------------------------------------
In addition to this change, would it make sense to have a flag to configure
DEFAULT_IN_MEMORY_ELEMENT_COUNT
(https://github.com/apache/beam/blob/release-2.34.0/sdks/java/core/src/main/java/org/apache/beam/sdk/transforms/join/CoGbkResult.java#L60)
similar to the setWorkerCacheMb flag?
> Use runtime information to improve CoGroupByKey caching
> -------------------------------------------------------
>
> Key: BEAM-13541
> URL: https://issues.apache.org/jira/browse/BEAM-13541
> Project: Beam
> Issue Type: Improvement
> Components: sdk-ideas
> Reporter: Sunil Pedapudi
> Assignee: Robert Bradshaw
> Priority: P2
> Time Spent: 40m
> Remaining Estimate: 0h
>
> Currently, CoGroupByKey creates UnionTables that are Flattened. The Flattened
> output is processed by a GroupByKey to produce a CoGbkResult (via
> ConstructCoGbkResultFn).
>
> Given the performance of CoGBK is greatly impacted based on the which
> elements are cached in the (finitely sized) in-memory results, it would be
> useful if CoGbkResult can use runtime information to prioritize which
> elements are stored in-memory.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
