[
https://issues.apache.org/jira/browse/SPARK-29881?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Devyn Cairns updated SPARK-29881:
---------------------------------
Description:
I have an interesting situation where I'm calling functions that are relatively
expensive from Spark SQL, and then using the result several times in a loop
through {{transform}}.
Although the WholeStageCodegen is usually helpful, it always calls expressions
as they're used, which means that in the case of, for example:
{{SELECT transform(sequence(0, 32), x -> expensive_result * x)}}
{{FROM (}}
{{ SELECT expensive_operation(foo) AS expensive_result FROM source}}
{{)}}
the expensive_operation function will almost certainly be called 32 times for
each source row, without any explicit way to cache that value intermediately.
I've found a workaround for now is to insert something like {{.filter \{ _ =>
true }}} in the middle, which will create a barrier to whole-stage codegen
without much negative impact, aside from preventing other optimizations like
PushDown. This does indeed produce the intended result and expensive_operation
is only run once.
But it would be great to have an API on Dataset like {{.barrier()}} to
introduce an explicit barrier to whole-stage codegen without adding any
additional behavior or getting in the way of any PushDown optimizations.
was:
I have an interesting situation where I'm calling functions that are relatively
expensive from Spark SQL, and then using the result several times in a loop
through {{transform}}.
Although the WholeStageCodegen is usually helpful, it always calls expressions
as they're used, which means that in the case of, for example:
{{SELECT transform(sequence(0, 32), x -> expensive_result * x)}}
{{FROM (}}
{{ SELECT expensive_operation(foo) AS expensive_result FROM source}}
{{)}}
the expensive_operation function will almost certainly be called 32 times for
each source row, without any explicit way to cache that value intermediately.
I've found a workaround for now is to insert something like {{.filter \{ _ =>
true }}} in the middle, which will create a barrier to whole-stage codegen
without much negative impact, aside from preventing other optimizations like
PushDown. This does indeed produce the intended result and expensive_operation
is only run once.
But it would be great to have an API on Dataset like {{.barrier()}} to
introduce an explicit barrier to whole-stage codegen without adding any
additional behavior or getting in the way of any PushDown optimizations.
> Introduce API for manually breaking up dataset plan
> ---------------------------------------------------
>
> Key: SPARK-29881
> URL: https://issues.apache.org/jira/browse/SPARK-29881
> Project: Spark
> Issue Type: Wish
> Components: SQL
> Affects Versions: 2.4.4
> Reporter: Devyn Cairns
> Priority: Trivial
>
> I have an interesting situation where I'm calling functions that are
> relatively expensive from Spark SQL, and then using the result several times
> in a loop through {{transform}}.
> Although the WholeStageCodegen is usually helpful, it always calls
> expressions as they're used, which means that in the case of, for example:
> {{SELECT transform(sequence(0, 32), x -> expensive_result * x)}}
> {{FROM (}}
> {{ SELECT expensive_operation(foo) AS expensive_result FROM source}}
> {{)}}
> the expensive_operation function will almost certainly be called 32 times for
> each source row, without any explicit way to cache that value intermediately.
> I've found a workaround for now is to insert something like {{.filter \{ _ =>
> true }}} in the middle, which will create a barrier to whole-stage codegen
> without much negative impact, aside from preventing other optimizations like
> PushDown. This does indeed produce the intended result and
> expensive_operation is only run once.
> But it would be great to have an API on Dataset like {{.barrier()}} to
> introduce an explicit barrier to whole-stage codegen without adding any
> additional behavior or getting in the way of any PushDown optimizations.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
