hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ning Zhang (JIRA)" <>
Subject [jira] Commented: (HIVE-1950) Block merge for RCFile
Date Wed, 16 Feb 2011 01:31:57 GMT


Ning Zhang commented on HIVE-1950:

2nd Round:
  same as my previous comment: revert the change if it belongs to another JIRA

  line 274, 310: you may not need the updateOnly variable in StatsWork. Instead you can just
check HiveConf.ConfVars.HIVE_STATS_ATOMIC. 
  missing Apache liense header
  the jobClose() functions should handle the exceptional case when abort is true (similar
to what FileSinkOperator does) or an exception was thrown from the hadoopo layber but it failed
to call close(abort=true). 
  also in jobClose(), the partition's old directory is first moved to backup directory and
then the intermediate directory is moved to the partition's destination directory. All this
is done when the partition is online (other queries can read the partition's directory). You
may want to create a follow-up JIRA to make this partition as offline during the move. 

> Block merge for RCFile
> ----------------------
>                 Key: HIVE-1950
>                 URL:
>             Project: Hive
>          Issue Type: New Feature
>            Reporter: He Yongqiang
>            Assignee: He Yongqiang
>         Attachments: HIVE-1950.1.patch, HIVE-1950.2.patch, HIVE-1950.3.patch, HIVE-1950.4.patch
> In our env, there are a lot of small files inside one partition/table. In order to reduce
the namenode load, we have one dedicated housekeeping job running to merge these file. Right
now the merge is an 'insert overwrite' in hive, and requires decompress the data and compress
it. This jira is to add a command in Hive to do the merge without decompress and recompress
the data.
> Something like "alter table tbl_name [partition ()] merge files". In this jira the new
command will only support RCFile, since there need some new APIs to the fileformat.

This message is automatically generated by JIRA.
For more information on JIRA, see:


View raw message