[jira] Commented: (HIVE-439) merge small files after a map-only job

["Zheng Shao (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-439?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12722513#action_12722513 ] 

Zheng Shao commented on HIVE-439:
---------------------------------

Testing. Will commit once the test succeeded.

A side note:
One more comment is the default value of hive.merge.size.per.mapper. In both HiveConf.java and conf/hive-default.xml, it's set to 10GB. That means most map-only jobs will run this optional task. Is that a bit too large? 

Given the usual block size of dfs 128MB, I think 100MB would be a good number. 100MB means that for identity select ("select * from"), we won't run this optional task. It also means that if we run the optional task, the output file size will be around 100MB which will most probably still be a single block.

Most problems in our world is caused by very small files, say 1MB or less. So I guess 100MB is probably good enough. 

What do you think?


> merge small files after a map-only job
> --------------------------------------
>
>                 Key: HIVE-439
>                 URL: https://issues.apache.org/jira/browse/HIVE-439
>             Project: Hadoop Hive
>          Issue Type: Improvement
>          Components: Query Processor
>    Affects Versions: 0.3.0, 0.3.1
>            Reporter: Namit Jain
>            Assignee: Namit Jain
>             Fix For: 0.4.0
>
>         Attachments: hive.439.1.patch, hive.439.2.patch, hive.439.3.patch, hive.439.4.patch
>
>
> There are cases when the input to a Hive job are thousands of small files. In this case, there is a mapper for each file. Most of the overhead for spawning all these mappers can be avoided if these small files are combined into fewer larger files.
> The problem can also be addressed by having a mapper span multiple blocks as in:
> https://issues.apache.org/jira/browse/HIVE-74
> Bit, it also makes sense in HIVE to merge files whenever possible.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.