[
https://issues.apache.org/jira/browse/HBASE-11900?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anoop Sam John updated HBASE-11900:
-----------------------------------
Component/s: mapreduce
> Optimization for incremental load reducer
> -----------------------------------------
>
> Key: HBASE-11900
> URL: https://issues.apache.org/jira/browse/HBASE-11900
> Project: HBase
> Issue Type: Improvement
> Components: HFile, mapreduce
> Affects Versions: 0.98.6
> Reporter: Yi Deng
> Priority: Minor
>
> In current implementation, the key of reducer configured by
> HFileOutputFormat.configureIncrementalLoad, is row. So, the reducer has to do
> an in-memory sort before writing key values to the disk. When we meet with
> some rows with a huge number of comlumns/versions, there could be OOM.
> A better way is:
> Use the KeyValue as the key, value can be a NullWritable. Partioner
> partitions the KeyValue only by it's row part. Set a sort comparator that
> sort KeyValue with KeyValue.COMPARATOR
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
