[
https://issues.apache.org/jira/browse/HADOOP-5589?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12699452#action_12699452
]
Chris Douglas commented on HADOOP-5589:
---------------------------------------
bq. The patch could be backwards-compatible if the bitset was written to the
stream as VLongs (essentially what had been implemented in HADOOP-5589-2.patch
This seems like a reasonable approach to me. As long as this can read old data,
it's OK if it writes records that cannot be read by the old implementation.
bq. is there still a need to support users who were storing tuples over 64
values but with incorrect results?
Such data is already corrupt, so I don't think anything can be done for it.
> TupleWritable: Lift implicit limit on the number of values that can be stored
> -----------------------------------------------------------------------------
>
> Key: HADOOP-5589
> URL: https://issues.apache.org/jira/browse/HADOOP-5589
> Project: Hadoop Core
> Issue Type: Improvement
> Components: mapred
> Affects Versions: 0.21.0
> Reporter: Jingkei Ly
> Assignee: Jingkei Ly
> Attachments: HADOOP-5589-1.patch, HADOOP-5589-2.patch,
> HADOOP-5589-3.patch
>
>
> TupleWritable uses an instance field of the primitive type, long, which I
> presume is so that it can quickly determine if a position has been written to
> in its array of Writables (by using bit-shifting operations on the long
> field). The problem with this is that it implies that there is a maximum
> limit of 64 values you can store in a TupleWritable.
> An example of a use-case where I think this would be a problem is if you had
> two MR jobs with over 64 reduces tasks and you wanted to join the outputs
> with CompositeInputFormat - this will probably cause unexpected results in
> the current scheme.
> At the very least, the 64-value limit should be documented in TupleWritable.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
