Andrew McHarg created SPARK-27560:
-------------------------------------
Summary: HashPartitioner uses Object.hashCode which is not seeded
Key: SPARK-27560
URL: https://issues.apache.org/jira/browse/SPARK-27560
Project: Spark
Issue Type: Bug
Components: Java API
Affects Versions: 2.4.0
Environment: Notebook is running spark v2.4.0 local[*]
Python 3.6.6 (default, Sep 6 2018, 13:10:03)
[GCC 4.2.1 Compatible Apple LLVM 9.1.0 (clang-902.0.39.2)] on darwin
I imagine this would reproduce on all operating systems and most versions of
spark though.
Reporter: Andrew McHarg
Forgive the quality of the bug report here, I am a pyspark user and not super
familiar with the internals of spark, yet it seems I have a strange corner case
with the HashPartitioner.
This may already be known but repartition with HashPartitioner seems to assign
everything the same partition if data that was partitioned by the same column
is only partially read (say one partition). I suppose it is obvious concequence
of Object.hashCode being deterministic but took some while to track down.
Steps to repro:
# Get dataframe with a bunch of uuids say 10000
# repartition(100, 'uuid_column')
# save to parquet
# read from parquet
# collect()[:100] then filter using pyspark.sql.functions isin (yes I know
this is bad and sampleBy should probably be used here)
# repartition(10, 'uuid_column')
# Resulting dataframe will have all of its data in one single partition
Jupyter notebook for the above:
https://gist.github.com/robo-hamburger/4752a40cb643318464e58ab66cf7d23e
I think an easy fix would be to seed the HashPartitioner like many hashtable
libraries do to avoid denial of service attacks. It also might be the case this
is obvious behavior for more experienced spark users :)
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
