The bug description is a little misleading: the actual issue is that
.py files are not handled correctly when distributed by YARN. They're
added to "spark.submit.pyFiles", which, when processed by context.py,
explicitly whitelists certain extensions (see PACKAGE_EXTENSIONS),
and that does not include .py files.
On top of that, archives were not handled at all! They made it to the
driver's python path, but never made it to executors, since the mechanism
used to propagate their location (spark.submit.pyFiles) only works on
the driver side.
So, instead, ignore "spark.submit.pyFiles" and just build PYTHONPATH
correctly for both driver and executors. Individual .py files are
placed in a subdirectory of the container's local dir in the cluster,
which is then added to the python path. Archives are added directly.
The change, as a side effect, ends up solving the symptom described
in the bug. The issue was not that the files were not being distributed,
but that they were never made visible to the python application
running under Spark.
Also included is a proper unit test for running python on YARN, which
broke in several different ways with the previous code.
A short walk around of the changes:
- SparkSubmit does not try to be smart about how YARN handles python
files anymore. It just passes down the configs to the YARN client
code.
- The YARN client distributes python files and archives differently,
placing the files in a subdirectory.
- The YARN client now sets PYTHONPATH for the processes it launches;
to properly handle different locations, it uses YARN's support for
embedding env variables, so to avoid YARN expanding those at the
wrong time, SparkConf is now propagated to the AM using a conf file
instead of command line options.
- Because the Client initialization code is a maze of implicit
dependencies, some code needed to be moved around to make sure
all needed state was available when the code ran.
- The pyspark tests in YarnClusterSuite now actually distribute and try
to use both a python file and an archive containing a different python
module. Also added a yarn-client tests for completeness.
- I cleaned up some of the code around distributing files to YARN, to
avoid adding more copied & pasted code to handle the new files being
distributed.
Author: Marcelo Vanzin <vanzin@cloudera.com>
Closes#6360 from vanzin/SPARK-5479 and squashes the following commits:
bcaf7e6 [Marcelo Vanzin] Feedback.
c47501f [Marcelo Vanzin] Fix yarn-client mode.
46b1d0c [Marcelo Vanzin] Merge branch 'master' into SPARK-5479
c743778 [Marcelo Vanzin] Only pyspark cares about python archives.
c8e5a82 [Marcelo Vanzin] Actually run pyspark in client mode.
705571d [Marcelo Vanzin] Move some code to the YARN module.
1dd4d0c [Marcelo Vanzin] Review feedback.
71ee736 [Marcelo Vanzin] Merge branch 'master' into SPARK-5479
220358b [Marcelo Vanzin] Scalastyle.
cdbb990 [Marcelo Vanzin] Merge branch 'master' into SPARK-5479
7fe3cd4 [Marcelo Vanzin] No need to distribute primary file to executors.
09045f1 [Marcelo Vanzin] Style.
943cbf4 [Marcelo Vanzin] [SPARK-5479] [yarn] Handle --py-files correctly in YARN.
38112905bc