[
https://issues.apache.org/jira/browse/HIVE-20330?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16693177#comment-16693177
]
Adam Szita commented on HIVE-20330:
-----------------------------------
Attached [^HIVE-20330.0.patch]
The change in this patch is that we're not just serializing and putting one
InputJobInfo into JobConf, but rather always append to a list (or create it on
the first occurrence) of InputJobInfo instances in it.
This ensures that if multiple tables serve as inputs in a job, Pig can retrieve
information for each of the tables, not just the last one added.
I've also discovered a bug in InputJobInfo.writeObject() where the
ObjectOutputStream was closed by mistake after writing partition information in
a compressed manner. Closing the compressed writer inevitably closed the OOS on
the context and prevented any other objects to be written into OOS - I had to
fix that because it prevented serializing InputJobInfo instances inside a list.
> HCatLoader cannot handle multiple InputJobInfo objects for a job with
> multiple inputs
> -------------------------------------------------------------------------------------
>
> Key: HIVE-20330
> URL: https://issues.apache.org/jira/browse/HIVE-20330
> Project: Hive
> Issue Type: Bug
> Components: HCatalog
> Reporter: Adam Szita
> Assignee: Adam Szita
> Priority: Major
> Attachments: HIVE-20330.0.patch
>
>
> While running performance tests on Pig (0.12 and 0.17) we've observed a huge
> performance drop in a workload that has multiple inputs from HCatLoader.
> The reason is that for a particular MR job with multiple Hive tables as
> input, Pig calls {{setLocation}} on each {{LoaderFunc (HCatLoader)}} instance
> but only one table's information (InputJobInfo instance) gets tracked in the
> JobConf. (This is under config key {{HCatConstants.HCAT_KEY_JOB_INFO}}).
> Any such call overwrites preexisting values, and thus only the last table's
> information will be considered when Pig calls {{getStatistics}} to calculate
> and estimate required reducer count.
> In cases when there are 2 input tables, 256GB and 1MB in size respectively,
> Pig will query the size information from HCat for both of them, but it will
> either see 1MB+1MB=2MB or 256GB+256GB=0.5TB depending on input order in the
> execution plan's DAG.
> It should of course see 256.00097GB in total and use 257 reducers by default
> accordingly.
> In unlucky cases this will be seen as 2MB and 1 reducer will have to struggle
> with the actual 256.00097GB...
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
