[PIG-2578] Multiple Store-commands mess up mapred.output.dir.

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.8.1, 0.9.2
• Fix Version/s: 0.10.0, 0.11
• Hadoop Flags: Reviewed
• Assignee: Daniel Dai
• Reporter: Mithun Radhakrishnan
• Created: 10/Mar/12 02:23
• Updated: 24/Aug/12 03:41
• Resolved: 14/Apr/12 08:03

Description

When one runs a pig-script with multiple storers, one sees the following:
1. When run as a script, Pig launches a single job.
2. PigOutputCommitter::setupJob() calls the underlyingOutputCommitter::setupJob(), once for each storer. But the mapred.output.dir is the same for both calls, even though the storers write to different locations.

This was originally seen in HCATALOG-276, when HCatalog's end-to-end tests are run against Pig.
(https://issues.apache.org/jira/browse/HCATALOG-276)

Sample pig-script (near identical to HCatalog's Pig_Checkin_4 test):

a = load 'keyvals' using org.apache.hcatalog.pig.HCatLoader();
split a into b if key<200, c if key >=200;
store b into 'keyvals_lt200' using org.apache.hcatalog.pig.HCatStorer();
store c into 'keyvals_ge200' using org.apache.hcatalog.pig.HCatStorer();

I've suggested a workaround in HCat for the time being, but I think this might be something that needs fixing in Pig.

Thanks.

Issue Links

• blocks: HCATALOG-276 After merging in HCATALOG-237 related changes Pig scripts with more than one store fail
• is related to: PIG-2870 pigServer.openIterator fails for jobs with no input splits