[
https://issues.apache.org/jira/browse/CRUNCH-528?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14561507#comment-14561507
]
Brandon Vargo commented on CRUNCH-528:
--------------------------------------
Identity instead of equality in cmp is a small speed optimization but shouldn't
affect correctness. If two objects are equal in the value sense but not
identical objects, then it'll fall through to either compareTo for Comparable
objects, which should return 0 for a well-formed comparison, or hashCode for
non-Comparable objects, which should also be the same for value equality. Not
calling equals is consistent with how TreeMap works internally, for example.
The common case is for two objects to not be equal, so calling compareTo and
looking for comparison == 0 is faster than calling equals and then having to
call compareTo anyways when equals returns false.
I don't have a better suggestion on how to handle joins with distinct objects
that have a hash collision. That's still broken unless the key implements
Comparable, since there is no secondary sort or equality check after the
hashCode, and TreeMap requires that compareTo be consistent with equals. This
isn't an issue when running on MapReduce because the shuffler requires that the
key be a WritableComparable and optionally uses an optimized
WritableComparator, but it is for in-memory mode, since it's just a TreeMap and
Pairs are allowed to contain any objects.
> Pair: Integer overflow during comparison can cause inconsistent sort.
> ---------------------------------------------------------------------
>
> Key: CRUNCH-528
> URL: https://issues.apache.org/jira/browse/CRUNCH-528
> Project: Crunch
> Issue Type: Bug
> Components: Core
> Reporter: Brandon Vargo
> Assignee: Josh Wills
> Priority: Minor
> Attachments: 0001-Pair-Fix-comparison-for-large-hash-codes.patch
>
>
> Pair uses the hash code of each value for comparison if the values are not
> themselves comparable. If the hash code values are too large, then the values
> will wrap when doing subtraction. This results in a comparison function that
> is not transitive.
> Among other things, this makes Joins using the in-memory pipeline not work,
> since the in-memory shuffler uses a TreeMap if the key type is Comparable.
> Since the key in a join is a Pair of the original key and a join tag, the key
> is always comparable. With a non-transitive comparison function, it is
> possible for the two join tags of the original key to sort differently,
> resulting in the two join tags not being adjacent for the original key. This
> results either in either the cross product erroneously producing no values in
> the case of an inner join, since the two join tags are not adjacent, or null
> values appearing when they should not in the case of an outer join.
> As a workaround, ensure that the key used in a Join is comparable.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
