[
https://issues.apache.org/jira/browse/SPARK-3030?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14120975#comment-14120975
]
Davies Liu commented on SPARK-3030:
-----------------------------------
I had update the PR to cleanup idle workers after 1 minute.
This feature can be turn of by "spark.python.worker.reuse=false". But I think
it's better to enable it by default, so we can test it.
We can disable it by default for the next release (if it's not stable enough).
> reuse python worker
> -------------------
>
> Key: SPARK-3030
> URL: https://issues.apache.org/jira/browse/SPARK-3030
> Project: Spark
> Issue Type: Improvement
> Components: PySpark
> Reporter: Davies Liu
> Assignee: Davies Liu
>
> Currently, it will fork an Python worker for each task, it will better if we
> can reuse the worker for later tasks.
> This will be very useful for large dataset with big broadcast, so it does not
> need to sending broadcast to worker again and again. Also it can reduce the
> overhead of launch a task.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
