github-actions[bot] commented on code in PR #65472:
URL: https://github.com/apache/doris/pull/65472#discussion_r3575691166
##########
fe/fe-core/src/main/java/org/apache/doris/nereids/rules/rewrite/PullUpJoinFromUnionAll.java:
##########
@@ -577,14 +577,14 @@ boolean isNotSupported(Plan plan) {
}
boolean comparePlan(Plan plan1, Plan plan2) {
+ if (plan1.getOutput().size() != plan2.getOutput().size()) {
Review Comment:
This guard is still too weak because same-table `LogicalOlapScan` nodes can
have the same output count while representing different scan identities. Two
production shapes reach this rule before nested-column pruning:
```text
UnionAll
Project(t.b)
Join(LogicalOlapScan(t INDEX mv_ab) -> [a, b], o1, on a = o1.k)
Project(t.c)
Join(LogicalOlapScan(t INDEX mv_ac) -> [a, c], o2, on a = o2.k)
```
`BindRelation` can create these direct selected-index scans before this
rewrite, and `computeOutput()` uses the selected index schema. Both scans have
the same `TableIdentifier` and arity, so this comparator maps outputs
positionally and can replace the second branch's `c` with the first branch's
`b`.
The same root cause applies even when outputs are identical but scan scope
differs, for example `LogicalOlapScan(t PARTITION p1)` vs `LogicalOlapScan(t
PARTITION p2)` or different manually selected tablets. After the comparator
treats them as the same common side, `constructNewJoin()` keeps only the first
scan as `newCommon`, so the second arm reads the wrong partition/tablet scope.
Please require full `LogicalOlapScan` identity before recording the
positional slot map, including selected index/output-column identity and
row-set restrictions such as selected/manual partitions and tablets, or reject
non-identical catalog-relation common sides.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
