neilconway commented on code in PR #22534:
URL: https://github.com/apache/datafusion/pull/22534#discussion_r3318522587
##########
datafusion/optimizer/src/eliminate_outer_join.rs:
##########
@@ -77,90 +104,152 @@ impl OptimizerRule for EliminateOuterJoin {
plan: LogicalPlan,
_config: &dyn OptimizerConfig,
) -> Result<Transformed<LogicalPlan>> {
- match plan {
- LogicalPlan::Filter(mut filter) => match
Arc::unwrap_or_clone(filter.input) {
+ let LogicalPlan::Filter(filter) = plan else {
+ return Ok(Transformed::no(plan));
+ };
+
+ // Descend through one or more Projection nodes until we find a Join.
+ // For each Projection we encounter, rewrite a working copy of the
+ // predicate by replacing references to projection output columns with
+ // the expressions that define them. Keep the filter's original
+ // predicate intact for eventual use in the rebuilt plan; the rewritten
+ // predicate is used only for the null-rejection analysis.
+ let mut rewritten_predicate = filter.predicate.clone();
+ let mut projections: Vec<Projection> = Vec::new();
+ let mut cur = Arc::clone(&filter.input);
+
+ let new_join = loop {
+ match cur.as_ref() {
+ LogicalPlan::Projection(p) => {
+ rewritten_predicate =
+ inline_through_projection(rewritten_predicate, p)?;
+ let next = Arc::clone(&p.input);
+ projections.push(p.clone());
Review Comment:
Yes, I spent some time noodling on that before sending the PR. Another
option would be wrapping the predicate in an `Option` and doing the clone
lazily. It seemed like overkill to me though, in the absence of some profiling
work to suggest this is actually a hotspot?
i.e., I can believe we definitely do way too many allocs in the optimizer,
but this is likely not the worst offender or anywhere close 😅
--
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]
