[
https://issues.apache.org/jira/browse/PIG-2014?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13025345#comment-13025345
]
Dmitriy V. Ryaboy commented on PIG-2014:
----------------------------------------
Proposal for a general-case fix:
add a @Nondeterministic annotation for UDFs, have the PushUpFilter check
whether the udf is deterministic or not when considering whether pushing up is
ok. Annotate the filter UDF that sample is rewritten to, accordingly.
> SAMPLE shouldn't be pushed up
> -----------------------------
>
> Key: PIG-2014
> URL: https://issues.apache.org/jira/browse/PIG-2014
> Project: Pig
> Issue Type: Bug
> Reporter: Jacob Perkins
>
> Consider the following code:
> {code:none}
> tfidf_all = LOAD '$TFIDF' AS (doc_id:chararray, token:chararray,
> weight:double);
> grouped = GROUP tfidf_all BY doc_id;
> vectors = FOREACH grouped GENERATE group AS doc_id, tfidf_all.(token,
> weight) AS vector;
> DUMP vectors;
> {code}
> This, of course, runs just fine. In a real example, tfidf_all contains
> 1,428,280 records. The reduce output records should be exactly the number of
> documents, which turn out to be 18,863 in this case. All well and good.
> The strangeness comes when you add a SAMPLE command:
> {code:none}
> sampled = SAMPLE vectors 0.0012;
> DUMP sampled;
> {code}
> Running this results in 1,513 reduce output records. The reduce output
> records be much much closer to 22 or 23 records (eg. 0.0012*18863).
> Evidently, Pig rewrites SAMPLE into filter, and then pushes that filter in
> front of the group. It shouldn't push that filter
> since the UDF is non-deterministic.
> Quick fix: If you add "-t PushUpFilter" to your command line when invoking
> pig this won't happen.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
