[
https://issues.apache.org/jira/browse/HBASE-14696?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ted Yu updated HBASE-14696:
---------------------------
Attachment: 14696-branch-1-v1.txt
> allowPartialResults in mapreduce Mappers
> ----------------------------------------
>
> Key: HBASE-14696
> URL: https://issues.apache.org/jira/browse/HBASE-14696
> Project: HBase
> Issue Type: Improvement
> Components: mapreduce
> Affects Versions: 2.0.0, 1.1.0
> Reporter: Mindaugas Kairys
> Assignee: Ted Yu
> Attachments: 14696-branch-1-v1.txt, 14696-v1.txt
>
>
> It is currently impossible to get partial results in mapreduce mapper jobs.
> When setting setAllowPartialResults(true) for scan jobs, they still fail with
> OOME on large rows.
> The reason is that Scan field allowPartialResults is lost during job creation:
> 1. User creates a Job and sets a scan object via
> TableMapReduceUtil.initTableMapperJob(table_name, scanObj,...) -> which puts
> a result of TableMapReduceUtil.convertScanToString(scanObj) to the job config.
> 2. When the job starts - method TableInputFormat.setConfig retrieves a scan
> string from config and converts it to Scan object by calling
> TableMapReduceUtil.convertStringToScan - which results in a Scan object with
> a field allowPartialResults always set to false.
> I have tried to experiment and modify a TableInputFormat method setConfig()
> by forcing all scans to allow partial results and after this all jobs
> succeeded with no more OOME and I also noticed that mappers began to get
> partial results (Result.isPartial()).
> My use case is very simple - I just have large rows and expect a mapper to
> get them partially - to get same rowid several times with different key/value
> records.
> This would allow me not to worry about implementing my own result
> partitioning solution, which i would encounter in case the big amount of
> result key values could be transparently returned for a single large row.
> And from the other side - if a Scan object can return several records for the
> same rowid (partial results), perhaps the mapper should do the same.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
