[
https://issues.apache.org/jira/browse/SPARK-26380?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Peiyu Zhuang updated SPARK-26380:
---------------------------------
Description:
We configure a shuffle manager that is compatible with the original Spark
shuffle manager (but with other network/storage implementation) in our
environment. It doesn't require any modification to Spark source code and It
works fine until we enable dynamic allocation. The problem is that in
{{ExternalShuffleBlockResolver}} class it checks the name of the shuffle
manager class and raise an {{UnsupportedOperationException}} if shuffle manager
is not {{SortShuffleManager}} or {{UnsafeShuffleManager}}.
Since shuffle manager is totally configurable, I think user should be able to
decide which shuffle manager to use when dynamic allocation is enabled.
Maybe we could change {{UnsupportedOperationException}} to a warning log
telling user that the shuffle manager they are using may not be compatible with
dynamic allocation and give them detail reasons why. On the other hand, user
may still choose to use their own shuffle manager while knowing the risk behind
it.
was:
We configure a shuffle manager that is compatible with the original Spark
shuffle manager (but with other network/storage implementation) in our
environment. It doesn't require any modification to Spark source code and It
works fine until we enable dynamic allocation. The problem is that in
{{ExternalShuffleBlockResolver}} class it checks the name of the shuffle
manager class and raise an {{UnsupportedOperationException}} if shuffle manager
is not {{SortShuffleManager}} or {{UnsafeShuffleManager}}.
Since shuffle manager is totally configurable, I think user should be able to
decide which shuffle manager to use when dynamic allocation is enabled.
Maybe we could change {{UnsupportedOperationException}} to a warning log
telling user that the shuffle manager they are using may not be compatible with
dynamic allocation and gives them detail reasons why. On the other hand, user
may still choose to use their own shuffle manager while knowing the risk behind
it.
> ExternalShuffleBlockResolver should not block shuffle managers by name.
> -----------------------------------------------------------------------
>
> Key: SPARK-26380
> URL: https://issues.apache.org/jira/browse/SPARK-26380
> Project: Spark
> Issue Type: Bug
> Components: Shuffle
> Affects Versions: 2.3.2
> Reporter: Peiyu Zhuang
> Priority: Major
>
> We configure a shuffle manager that is compatible with the original Spark
> shuffle manager (but with other network/storage implementation) in our
> environment. It doesn't require any modification to Spark source code and It
> works fine until we enable dynamic allocation. The problem is that in
> {{ExternalShuffleBlockResolver}} class it checks the name of the shuffle
> manager class and raise an {{UnsupportedOperationException}} if shuffle
> manager is not {{SortShuffleManager}} or {{UnsafeShuffleManager}}.
> Since shuffle manager is totally configurable, I think user should be able to
> decide which shuffle manager to use when dynamic allocation is enabled.
> Maybe we could change {{UnsupportedOperationException}} to a warning log
> telling user that the shuffle manager they are using may not be compatible
> with dynamic allocation and give them detail reasons why. On the other hand,
> user may still choose to use their own shuffle manager while knowing the risk
> behind it.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
