[
https://issues.apache.org/jira/browse/SPARK-18161?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15614752#comment-15614752
]
Sloane Simmons commented on SPARK-18161:
----------------------------------------
I changed the importance from minor to major since there is a definite impact
to this to PySpark users ( programs serializing large objects crash without
this change ), but there is a workaround ( create a SparkContext with a
serializer that uses {{pickle.HIGHEST_PROTOCOL}} and monkey-patch
Broadcast#dumps to do the same ).
> Default PickleSerializer pickle protocol doesn't handle > 4GB objects
> ---------------------------------------------------------------------
>
> Key: SPARK-18161
> URL: https://issues.apache.org/jira/browse/SPARK-18161
> Project: Spark
> Issue Type: Improvement
> Components: PySpark
> Affects Versions: 2.0.0, 2.0.1
> Reporter: Sloane Simmons
>
> When broadcasting a fairly large numpy matrix in a Spark 2.0.1 program, there
> is an error serializing the object with:
> {{OverflowError: cannot serialize a bytes object larger than 4 GiB}}
> in the stack trace.
> This is because Python's pickle serialization (with protocol <= 3) uses a
> 32-bit integer for the object size, and so cannot handle objects larger than
> 4 gigabytes. This was changed in Protocol 4 of pickle
> (https://www.python.org/dev/peps/pep-3154/#bit-opcodes-for-large-objects) and
> is available in Python 3.4+.
> I would like to use this protocol for broadcasting and in the default
> PickleSerializer where available to make pyspark more robust to broadcasting
> large variables.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
