[
https://issues.apache.org/jira/browse/SPARK-26412?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16836480#comment-16836480
]
Weichen Xu commented on SPARK-26412:
------------------------------------
Discuss with [~mengxr] , discard proposal (2), this should be something like
(user perspective):
{code:java}
def mapper_func(batch_iter):
# do some initialization
for batch in batch_iter:
# do some computation on `batch` (a pandas dataframe)
# and generate `result` (also a pandas dataframe)
yield result
df.mapPartitions(mapper_func){code}
> Allow Pandas UDF to take an iterator of pd.DataFrames or Arrow batches
> ----------------------------------------------------------------------
>
> Key: SPARK-26412
> URL: https://issues.apache.org/jira/browse/SPARK-26412
> Project: Spark
> Issue Type: New Feature
> Components: PySpark
> Affects Versions: 3.0.0
> Reporter: Xiangrui Meng
> Assignee: Weichen Xu
> Priority: Major
>
> Pandas UDF is the ideal connection between PySpark and DL model inference
> workload. However, user needs to load the model file first to make
> predictions. It is common to see models of size ~100MB or bigger. If the
> Pandas UDF execution is limited to batch scope, user need to repeatedly load
> the same model for every batch in the same python worker process, which is
> inefficient. I created this JIRA to discuss possible solutions.
> Essentially we need to support "start()" and "finish()" besides "apply". We
> can either provide those interfaces or simply provide users the iterator of
> batches in pd.DataFrame or Arrow table and let user code handle it.
> Another benefit is with iterator interface and asyncio from Python, it is
> flexible for users to implement data pipelining.
> cc: [~icexelloss] [~bryanc] [~holdenk] [~hyukjin.kwon] [~ueshin] [~smilegator]
--
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
