Tim Armstrong has posted comments on this change.

Change subject: IMPALA-1286: Extract common conjuncts from disjunctions.
......................................................................


Patch Set 2: Code-Review+1

(4 comments)

http://gerrit.cloudera.org:8080/#/c/4877/1/fe/src/main/java/org/apache/impala/rewrite/ExtractCommonConjunctRule.java
File fe/src/main/java/org/apache/impala/rewrite/ExtractCommonConjunctRule.java:

Line 31:  * Extracts common conjuncts from a single disjunctive 
CompoundPredicate.
> Expanded comments on ExprRewriter and modified comment here. 
I agree we don't need to document the framework here, but I think it's 
non-obvious that implementing it for a single CompoundPredicate is equally as 
powerful as implementing it for n CompoundPredicates.

If I understand correctly it relies on the bottom-up application of rewrite 
rules to "bubble up" common predicates.

E.g. "This rule extracts common conjuncts from multiple disjunctions when it is 
applied recursively bottom-up to a tree of CompoundPredicates".


Line 49: 
> Agree that this is problematic. Unfortunately, also non-trivial to fix and 
Codegen in the BE generally shouldn't be N^2, although the constant factor is 
much higher.


http://gerrit.cloudera.org:8080/#/c/4877/2/fe/src/main/java/org/apache/impala/rewrite/ExtractCommonConjunctRule.java
File fe/src/main/java/org/apache/impala/rewrite/ExtractCommonConjunctRule.java:

Line 51:     List<Expr> child0Conjuncts = expr.getChild(0).getConjuncts();
> checkstate that the child conjuncts aren't empty?
true or x => true seems like a normalisation rule we should add in a follow-up 
patch


http://gerrit.cloudera.org:8080/#/c/4877/1/fe/src/test/java/org/apache/impala/analysis/ExprRewriteRulesTest.java
File fe/src/test/java/org/apache/impala/analysis/ExprRewriteRulesTest.java:

Line 146:         "(!(int_col=5 or tinyint_col > 9) and double_col = 7) or " +
> Completely agree about having additional rules for normalization. I view th
I think we should just be mindful that this change without normalisation added 
a performance cliff where e.g. flipping x < y to y > x can change the plan. I 
think prior to this change the two were equivalent.

Seems ok as an intermediate state but we shouldn't consider the feature 
finished until we've made it a little more robust. I think this probably also 
relates to the constant expr -> literal rewrite that's on my plate.


-- 
To view, visit http://gerrit.cloudera.org:8080/4877
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-MessageType: comment
Gerrit-Change-Id: I3cf9b950afaa3fd753d1b09ba5e540b5258940ad
Gerrit-PatchSet: 2
Gerrit-Project: Impala-ASF
Gerrit-Branch: master
Gerrit-Owner: Alex Behm <[email protected]>
Gerrit-Reviewer: Alex Behm <[email protected]>
Gerrit-Reviewer: Bharath Vissapragada <[email protected]>
Gerrit-Reviewer: Marcel Kornacker <[email protected]>
Gerrit-Reviewer: Tim Armstrong <[email protected]>
Gerrit-HasComments: Yes

Reply via email to