[CRUNCH-557] Fix file distribution from HDFS in Crunch-on-Spark

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Assignee: Unassigned
• Reporter: Josh Wills
• Created: 02/Sep/15 22:54
• Updated: 25/Sep/15 03:09

Description

From the user list:

I was trying to determine effect of changing JoinStrategy on a Spark pipeline. I noticed that my pipeline works fine with DefaultJoinStrategy, however I could not get it to working with MapSideJoinStrategy and BloomFilterJoinStrategy. For MapSideJoinStrategy I get an exceptions[1] on driver itself and for BloomFilterJoinStrategy I get exceptions[2] in one of the stages. I have not tried to do any configuration changes but I did run tests with datasets of different sizes to ensure that my PCollection is small enough to fit in memory. I am running spark in yarn-client mode with Crunch 0.11.0-cdh5.4.2.

[1] https://gist.github.com/anonymous/15d6c691b743ad392d42
[2] https://gist.github.com/anonymous/b02a82401a30a69f1cff

The bug is in the SparkRuntime.distributeFiles method, which needs to include a scheme for the URI it's handing to Spark.