Critical Description
Please find below a minimal example of a Pig script that uses splits and replicated joins and where the output differs between MapReduce and TEZ as execution engine. The attachment also contains the sample input data.
The expected output, as created by MapReduce engine is:
(id1,123,A,) (id2,234,,B) (id3,456,,) (id4,567,A,)
whereas TEZ produces
(id1,123,A,A) (id2,234,B,B) (id3,456,,) (id4,567,A,A)
Removing the USING 'replicated' and using a regular join yields correct results. I am not sure if this is a Pig issue or a TEZ issue. However, as this issue silently can lead to data corruption I rated it critical. So far searching didn't indicate a similar bug or anybody being aware of it.
classdata = LOAD '/tez_bug_input1.csv' USING PigStorage(',') AS (classid:chararray, class:chararray); data = LOAD '/tez_bug_input2.csv' USING PigStorage(',') AS (eventid:chararray, classid:chararray); basedata = LOAD '/tez_bug_input3.csv' USING PigStorage(',') AS (eventid:chararray, foo:int); dataJclassdata = JOIN classdata BY classid, data BY classid; SPLIT dataJclassdata INTO classA IF class == 'A', classB IF class == 'B'; dataA = JOIN basedata BY eventid LEFT OUTER, classA BY data::eventid USING 'replicated'; dataA = foreach dataA generate basedata::eventid as eventid , basedata::foo as foo , classA::classdata::class as classA; dataB = JOIN dataA BY eventid LEFT OUTER, classB BY eventid USING 'replicated'; dataB = foreach dataB generate dataA::eventid as eventid , dataA::foo as foo , dataA::classA as classA , classB::classdata::class as classB; DUMP dataB;
