gengliangwang commented on a change in pull request #24598: [SPARK-27699][SQL]
Partially push down disjunctive predicated in Parquet/ORC
URL: https://github.com/apache/spark/pull/24598#discussion_r284119542
##########
File path:
sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetFilters.scala
##########
@@ -527,11 +527,22 @@ private[parquet] class ParquetFilters(
}
case sources.Or(lhs, rhs) =>
+ // The Or predicate is convertible when both of its children can be
pushed down.
+ // That is to say, if one/both of the children can be partially pushed
down, the Or
+ // predicate can be partially pushed down as well.
+ //
+ // Here is an example used to explain the reason.
+ // Let's say we have
+ // (a1 AND a2) OR (b1 AND b2),
+ // a1 and b1 is convertible, while a2 and b2 is not.
+ // The predicate can be converted as
+ // (a1 OR b1) AND (a1 OR b2) AND (a2 OR b1) AND (a2 OR b2)
Review comment:
> The rule is correct, but it looks a little dangerous to set true blindly
without considering the given parameter canPartialPushDownConjuncts. Could you
add a test case for that complex example like OR under NOT in the deep
predicate tree?
@dongjoon-hyun nice catch. However, here the `Not` predicate won't have a
child as `Or` or `And` predicate because of `BooleanSimplification`
optimization rule.
I have updated the code, but a new test case seems unnecessary. `Not(a Or
b)` can be converted as `Not(a) And Not(b)`
Also, I have created a PR for pushing down `Not` operator before(for double
insurance), but seems the PR made thing too complex:
https://github.com/apache/spark/pull/22687
> Also, it would be great to add more higher level test case in
SQLQuerySuite.scala to show the benefit of this additional predicate pushdown
a1 OR b1. Could you add that, too?
How can we verify the predicate is pushed down? Match the `OrcScan` and
check the `pushedFilters`? Only Orc V2 can be checked in this way currently.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
