[
https://issues.apache.org/jira/browse/SPARK-10710?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Josh Rosen updated SPARK-10710:
-------------------------------
Description:
The {{spark.shuffle.spill=false}} configuration doesn't make much sense
nowadays: I think that this configuration was only added as an escape-hatch to
guard against bugs when spilling was first added. Similarly, setting
{{spark.sql.planner.externalSort=false}} doesn't make sense in newer releases:
many new implementations, such as Tungsten, completely ignore this flag, so
it's not applied in a consistent way.
In order to reduce complexity, I think that we should remove the ability to
disable spilling. Note that the {{tungsten-shuffle}} manager already does not
respect this setting, so removing this configuration is a blocker to being able
to unify the two sort-shuffle implementations.
was:
The {{spark.shuffle.spill=false}} configuration doesn't make much sense
nowadays: I think that this configuration was only added as an escape-hatch to
guard against bugs when spilling was first added.
In order to reduce complexity, I think that we should remove the ability to
disable spilling. Note that the {{tungsten-shuffle}} manager already does not
respect this setting, so removing this configuration is a blocker to being able
to unify the two sort-shuffle implementations.
> Remove ability to set spark.shuffle.spill=false and
> spark.sql.externalSort=false
> --------------------------------------------------------------------------------
>
> Key: SPARK-10710
> URL: https://issues.apache.org/jira/browse/SPARK-10710
> Project: Spark
> Issue Type: Sub-task
> Components: Shuffle
> Reporter: Josh Rosen
> Assignee: Josh Rosen
>
> The {{spark.shuffle.spill=false}} configuration doesn't make much sense
> nowadays: I think that this configuration was only added as an escape-hatch
> to guard against bugs when spilling was first added. Similarly, setting
> {{spark.sql.planner.externalSort=false}} doesn't make sense in newer
> releases: many new implementations, such as Tungsten, completely ignore this
> flag, so it's not applied in a consistent way.
> In order to reduce complexity, I think that we should remove the ability to
> disable spilling. Note that the {{tungsten-shuffle}} manager already does not
> respect this setting, so removing this configuration is a blocker to being
> able to unify the two sort-shuffle implementations.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
