[
https://issues.apache.org/jira/browse/HIVE-1343?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12867557#action_12867557
]
Ning Zhang commented on HIVE-1343:
----------------------------------
Yongqiang this patch only exposes the FileInputReader to the client and the
client has to merge the file locally. This won't be scalable. What we should do
is to run this merge job as a map-only job so that it can be run in parallel.
Talked with Dhruba and he think it would be possible to make it a map-only job.
The idea is to define a new RecordReader that does not do decompression and
iterate over records. Instead it iterates over uncompressed blocks.
> add an interface in RCFile to support concatenation of two files without
> (de)compression
> ----------------------------------------------------------------------------------------
>
> Key: HIVE-1343
> URL: https://issues.apache.org/jira/browse/HIVE-1343
> Project: Hadoop Hive
> Issue Type: New Feature
> Affects Versions: 0.6.0
> Reporter: Ning Zhang
> Assignee: He Yongqiang
> Fix For: 0.6.0
>
> Attachments: HIVE-1343.1.patch
>
>
> If two files are concatenated, we need to read each record in these files and
> write them back to the destination file. The IO cost is mostly unavoidable
> due to the lack of append functionality in HDFS. However the CPU cost could
> be significantly reduced by avoiding compression and decompression of the
> files.
> The File Format layer should provide API that implement the block-level
> concatenation.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
