You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@spark.apache.org by "Tengfei Huang (Jira)" <ji...@apache.org> on 2025/09/15 03:50:00 UTC
[jira] [Updated] (SPARK-53575) Retry entire consumer stages for ShuffleFetchFailed when the producer stage is non-deterministic
[ https://issues.apache.org/jira/browse/SPARK-53575?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Tengfei Huang updated SPARK-53575:
----------------------------------
Description:
With https://issues.apache.org/jira/browse/SPARK-51756 we have computed a checksum for shuffle output which can be used to detect shuffle output changes while retrying.
Once ShuffleFetchFailed happens and checksum mismatch detected, we need to retry the entire consumer stages since the producer stage generates different data across different attempts.
In this case, retrying the failed tasks for the consumer stages is not sufficient, as the data may have been changed for other consumer tasks by the non-deterministic producers.
was:
Retry the entire consumer stages during ShuffleFetchFailed, in the case that the producer stage generates different data across different attempts.
In this case, retrying the failed tasks for the consumer stages is not sufficient, as the data may have been changed for other consumer tasks by the non-deterministic producers.
> Retry entire consumer stages for ShuffleFetchFailed when the producer stage is non-deterministic
> ------------------------------------------------------------------------------------------------
>
> Key: SPARK-53575
> URL: https://issues.apache.org/jira/browse/SPARK-53575
> Project: Spark
> Issue Type: Improvement
> Components: Spark Core
> Affects Versions: 4.0.1
> Reporter: Tengfei Huang
> Priority: Major
>
> With https://issues.apache.org/jira/browse/SPARK-51756 we have computed a checksum for shuffle output which can be used to detect shuffle output changes while retrying.
> Once ShuffleFetchFailed happens and checksum mismatch detected, we need to retry the entire consumer stages since the producer stage generates different data across different attempts.
> In this case, retrying the failed tasks for the consumer stages is not sufficient, as the data may have been changed for other consumer tasks by the non-deterministic producers.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org