[
https://issues.apache.org/jira/browse/PIG-250?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Olga Natkovich updated PIG-250:
-------------------------------
Attachment: PIG-250_v2.patch
Arun helped to diagnose the problem. It turned out that we were not passing the
right directory to the task to write to and as the result all tasks (primary +
speculative) were writing to the same place overwriting each others output.
The attached patch fixes the issue and also reverses the prior change of
disabling speculative execution. With this changes, by default the speculative
execution will be consistent with hadoop default (currently on) and will be
controlled by configuration (hadoop-site.xml)
I tested that speculative execution works now and that I can control it from
hadoop-site.xml.
Thanks, Arun.
One of the committers - please review.
> Pig is broken with speculative execution
> ----------------------------------------
>
> Key: PIG-250
> URL: https://issues.apache.org/jira/browse/PIG-250
> Project: Pig
> Issue Type: Bug
> Reporter: Olga Natkovich
> Attachments: PIG-250.patch, PIG-250_v2.patch
>
>
> If I have speculative execution turned on, the following script fails:
> a = load 'studenttab20m' as (name, age, gpa);
> b = load 'votertab10k' as (name, age, registration, contributions);
> c = filter a by age < '50';
> d = filter b by age < '50';
> e = cogroup c by (name, age), d by (name, age) parallel 10;
> f = foreach e generate flatten(c), flatten(d) parallel 10;
> g = group f by registration parallel 10;
> h = foreach g generate group, SUM(f.d::contributions) parallel 10;
> i = order h by ($1, $0);
> store i into 'out';
> I traced this to the fact that the first MR job produces one or more empty
> outputs from the reducer. This happened on the reducers that happened to have
> second task running.
> I am not sure what the issue is and I am working with hadoop guys to
> investigate. Until this issue is resolved, I would like to trun speculative
> execution off.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
