[
https://issues.apache.org/jira/browse/SPARK-26858?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16772570#comment-16772570
]
Hyukjin Kwon edited comment on SPARK-26858 at 2/20/19 3:00 AM:
---------------------------------------------------------------
Yes .. that's matched with what I was thinking. Thanks for clarification,
[~bryanc].
The discussion is getting long .. to make it short ..
- Arrow requires to know its schema for its output
- Schema can be only known after R native function execution for its output
- It might be able to work around by, for instance, set the DataFrame's
output as {{binary}} type for its output, and then construct Arrow streaming
format by, somehow, lazily providing the schema (for its batch or {{Schema}}).
Given my try, it's quite hacky since it basically means you need to send
{{Schema}} from executor to driver (to be used in {{collect()}} later).
(One other possibility I was thinking about batches without {{Schema}} is that
we just send Arrow batch by Arrow batch, deserialize each, and then construct
an Arrow table, which is pretty different from Python side and hacky)
Yes, it's currently se/deserialized row by row bytes over socket.
was (Author: hyukjin.kwon):
Yes .. that's matched with what I was thinking. Thanks for clarification,
[~bryanc].
The discussion is getting long .. to make it short ..
- Arrow requires to know its schema for its output
- Schema can be only known after R native function execution for its output
- It might be able to work around by, for instance, set the DataFrame's
output as {{binary}} type for its output, and then construct Arrow streaming
format by, somehow, lazily providing the schema (for its batch or {{Schema}}).
Given my try, it's quite hacky since it basically means you need to send
{{Schema}} from executor to driver (to be used in {{collect()}} later).
Yes, it's currently se/deserialized row by row bytes over socket.
> Vectorized gapplyCollect, Arrow optimization in native R function execution
> ---------------------------------------------------------------------------
>
> Key: SPARK-26858
> URL: https://issues.apache.org/jira/browse/SPARK-26858
> Project: Spark
> Issue Type: Sub-task
> Components: SparkR, SQL
> Affects Versions: 3.0.0
> Reporter: Hyukjin Kwon
> Assignee: Hyukjin Kwon
> Priority: Major
>
> Unlike gapply, gapplyCollect requires additional ser/de steps because it can
> omit the schema, and Spark SQL doesn't know the return type before actually
> execution happens.
> In original code path, it's done via using binary schema. Once gapply is done
> (SPARK-26761). we can mimic this approach in vectorized gapply to support
> gapplyCollect.
--
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
