[
https://issues.apache.org/jira/browse/MAHOUT-320?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12841181#action_12841181
]
Sean Owen commented on MAHOUT-320:
----------------------------------
After the fact here, but I'm going to fix a number of small issues in
IntPairWritable, and three that I think are important enough to call out. I
keep seeing these kinds of subtle bugs:
- hashCode() and equals() not only must both be implemented, but must be
consistent. Frequency uses 'pair' in hashCode() but not in equals()
- and you can hash doubles much faster with RandomUtils.hashDouble()
- compareTo() must be consistent with equals(). It returns 0 exactly when
equals() returns true. Frequency.compareTo() can never return 0, which is never
correct, even if only because this breaks symmetry
- compareTo() that decides ordering based on two integer values should never be
implemented with a subtraction, unless you're absolutely certain the following
can't occur: trying to order a=Integer.MAX_VALUE and b=Integer.MIN_VALUE will
put a before b, if you base it on a-b's value, since it overflows to negative.
> Modify IntPairWritable in LDA implementation to be binary comparable to
> improve performance.
> --------------------------------------------------------------------------------------------
>
> Key: MAHOUT-320
> URL: https://issues.apache.org/jira/browse/MAHOUT-320
> Project: Mahout
> Issue Type: Improvement
> Components: Clustering
> Affects Versions: 0.3
> Reporter: Drew Farris
> Assignee: Robin Anil
> Priority: Minor
> Attachments: MAHOUT-320.patch, MAHOUT-320.patch, MAHOUT-320.patch,
> MAHOUT-320.patch, MAHOUT-320.patch
>
>
> Per discussion with Robin, modifying o.a.m.clustering.lda.IntPairWritable to
> be binary comparable will improve the performance of the comparison
> operations during a sort because no marshaling will need to occur to compare
> IntPairWritable instances.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
