[
https://issues.apache.org/jira/browse/SPARK-17995?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15586972#comment-15586972
]
Ryan Blue commented on SPARK-17995:
-----------------------------------
[~cloud_fan] and [~yhuai], I'd like to help fix this, but I'm not sure the best
way.
I started to write an analyzer rule that uses transformUp on the initial
logical plan, before unresovled aliases are resolved. That rule would find
outer joins and generate a map of attributes to replace to the new attribute
above the join, with the schema updated to be nullable and with a new exprId.
The attributes to replace come from the output of the outer join.
Where I ran into trouble was in replacing the attributes in the logical plan
above the a join. I don't think it is a good idea to have cases in the rule for
every possible plan, so I think we need a method to substitute attributes that
is implemented by nodes in the plan. That sounds like a larger patch than I
originally thought, so I wanted to make sure I'm going down the right path
before I put up a PR for it. What do you think?
> Use new attributes for columns from outer joins
> -----------------------------------------------
>
> Key: SPARK-17995
> URL: https://issues.apache.org/jira/browse/SPARK-17995
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 1.6.2, 2.0.0, 2.1.0
> Reporter: Ryan Blue
>
> Plans involving outer joins use the same attribute reference (by exprId) to
> reference columns above the join and below the join. This is a false
> equivalence that leads to bugs like SPARK-16181, in which an attributes were
> incorrectly replaced by the optimizer. The column has a different schema
> above the outer join because its values may be null. The fix for that issue,
> [PR #13884](https://github.com/apache/spark/pull/13884) has a TODO comment
> from [~cloud_fan] to fix this by using different attributes instead of
> needing to special-case outer joins in rules and this issue is to track that
> improvement.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
