[
https://issues.apache.org/jira/browse/FLINK-28889?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Weijie Guo updated FLINK-28889:
-------------------------------
Description: Hybrid shuffle does not support multiple consumer for single
subpartition data. This will bring some defects, such as the inability to
support partition reuse, speculative execution. In particular, it cannot
support broadcast optimization, that is, hybrid shuffle writes multiple copies
of broadcast data, This will cause a waste of memory and disk space and affect
the performance of shuffle write phase. Ideally, for the full spilling
strategy, any broadcast data (record or event) should only write one piece of
data in the memory, and the same is true for the disk. (was: Hybrid shuffle
writes multiple copies of broadcast data, This will cause a waste of memory and
disk space and affect the performance of shuffle write phase. Ideally, for the
full spilling strategy, any broadcast data (record or event) should only write
one piece of data in the memory, and the same is true for the disk. For
selective spilling strategy, if the broadcast edge is encountered, we should
consider directly turning it into the edge of HYBRID_FULL, or introducing
configuration option to decide whether to do this switch. )
> Hybrid shuffle writes multiple copies of broadcast data
> -------------------------------------------------------
>
> Key: FLINK-28889
> URL: https://issues.apache.org/jira/browse/FLINK-28889
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Network
> Affects Versions: 1.16.0
> Reporter: Weijie Guo
> Assignee: Weijie Guo
> Priority: Critical
> Fix For: 1.17.0
>
>
> Hybrid shuffle does not support multiple consumer for single subpartition
> data. This will bring some defects, such as the inability to support
> partition reuse, speculative execution. In particular, it cannot support
> broadcast optimization, that is, hybrid shuffle writes multiple copies of
> broadcast data, This will cause a waste of memory and disk space and affect
> the performance of shuffle write phase. Ideally, for the full spilling
> strategy, any broadcast data (record or event) should only write one piece of
> data in the memory, and the same is true for the disk.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
