[
https://issues.apache.org/jira/browse/SPARK-22239?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16452977#comment-16452977
]
Li Jin commented on SPARK-22239:
--------------------------------
[~hvanhovell], I have done a bit further research of UDF over rolling windows
and posted my results here:
[https://docs.google.com/document/d/14EjeY5z4-NC27-SmIP9CsMPCANeTcvxN44a7SIJtZPc/edit?usp=sharing]
TL; DR I think we can implement efficiently by computing window indices in the
JVM and pass the indices along with the window Python and do rolling over the
indices in Python. I have not addressed the issue of splitting the window
partition into smaller batches but I think it's doable as well. Would you be
interested in taking a look and let me know what you think?
> User-defined window functions with pandas udf
> ---------------------------------------------
>
> Key: SPARK-22239
> URL: https://issues.apache.org/jira/browse/SPARK-22239
> Project: Spark
> Issue Type: Sub-task
> Components: PySpark
> Affects Versions: 2.2.0
> Environment:
> Reporter: Li Jin
> Priority: Major
>
> Window function is another place we can benefit from vectored udf and add
> another useful function to the pandas_udf suite.
> Example usage (preliminary):
> {code:java}
> w = Window.partitionBy('id').orderBy('time').rangeBetween(-200, 0)
> @pandas_udf(DoubleType())
> def ema(v1):
> return v1.ewm(alpha=0.5).mean().iloc[-1]
> df.withColumn('v1_ema', ema(df.v1).over(window))
> {code}
--
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
