Stop copying LogicalPlan and Exprs in CommonSubexprEliminate (2-3% planning speed improvement)
#10835
Conversation
github-actions
bot
added
logical-expr
| @@ -870,37 +870,7 @@ impl LogicalPlan { | |||
| LogicalPlan::Filter { .. } => { | |||
| assert_eq!(1, expr.len()); | |||
| let predicate = expr.pop().unwrap(); | |||
|
|
|||
| // filter predicates should not contain aliased expressions so we remove any aliases | |||
There was a problem hiding this comment.
This code is extract into Filter::remove_aliases as it is needed for CSE.
| @@ -173,9 +179,8 @@ impl CommonSubexprEliminate { | |||
| &mut common_exprs, | |||
| )?; | |||
|
|
|||
| let mut new_input = self | |||
| .try_optimize(input, config)? | |||
| .unwrap_or_else(|| input.clone()); | |||
There was a problem hiding this comment.
The whole point of this PR is to remove clones -- I'll try and point out example places such as this where the plans and/or expressions were being cloned
| // WindowAggr: windowExpr=[[sum(c9) ORDER BY [c3 + c4] RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW]] | ||
| // --WindowAggr: windowExpr=[[sum(c9) ORDER BY [c3 + c4] RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW]] | ||
| // where, it is referred once by each `WindowAggr` (total of 2) in the plan. | ||
| let mut plan = LogicalPlan::Window(window.clone()); |
| let Window { | ||
| input, window_expr, .. | ||
| } = window; | ||
| plan = input.as_ref().clone(); |
| .zip(arrays_list.iter()) | ||
| .map(|(exprs, arrays)| { | ||
| exprs | ||
| .iter() | ||
| .cloned() |
| .iter() | ||
| .zip(aggr_expr.iter()) | ||
| .map(|(new_expr, old_expr)| { | ||
| new_expr.clone().alias_if_changed(old_expr.display_name()?) |
|
Update -- q1 gets significantly worse -- I'll try and profile it and figure out why |
|
UPDATE: I ran the wrong code. I am now rerunning... |
|
Ok, update here is that after some additional tracking of when plans are rewritten, this PR now improves planning performance by 2-3% so I think it is ready for review |
alamb
changed the title
CommonSubexprEliminateCommonSubexprEliminate (2-3% planning speed improvement)
|
I believe this PR has conflicts with #10939 which I am now resolving Thank you @crepererum for your review |
|
@peter-toth would you like to review this PR as well? |
There was a problem hiding this comment.
Thanks for pinging me @alamb, LGTM, only minor suggestions.
| /// Decimal128(Some(69999999999999),30,15) | ||
| /// AND lineitem.l_quantity < Decimal128(Some(2400),15,2) | ||
| /// ``` | ||
| pub fn remove_aliases(predicate: Expr) -> Result<Transformed<Expr>> { |
There was a problem hiding this comment.
I wonder if it make sense to move this method into Expr?
There was a problem hiding this comment.
I think it does -- I will do so as a follow on PR
|
|
||
| if !common_exprs.is_empty() { | ||
| new_input = build_common_expr_project_plan(new_input, common_exprs)?; | ||
| transformed = true; |
There was a problem hiding this comment.
I'm not sure we need this because if !common_exprs.is_empty() is true then rewrite_exprs.transformed must also be true a bit above.
There was a problem hiding this comment.
that is a good invariant -- I did not know that. Changed to an assert in in 8dd6256
| Ok(Transformed::new(expr, false, TreeNodeRecursion::Jump)) | ||
| } | ||
| Expr::Alias(_) => Ok(Transformed::new( | ||
| expr.unalias(), |
There was a problem hiding this comment.
Since we know that expr is an Expr::Alias maybe we could just use Expr::Alias(alias) => *alias.expr instead of calling .unalias(), that matches the expr again.
| /// Decimal128(Some(69999999999999),30,15) | ||
| /// AND lineitem.l_quantity < Decimal128(Some(2400),15,2) | ||
| /// ``` | ||
| pub fn remove_aliases(predicate: Expr) -> Result<Transformed<Expr>> { |
There was a problem hiding this comment.
I think it does -- I will do so as a follow on PR
|
|
||
| if !common_exprs.is_empty() { | ||
| new_input = build_common_expr_project_plan(new_input, common_exprs)?; | ||
| transformed = true; |
There was a problem hiding this comment.
that is a good invariant -- I did not know that. Changed to an assert in in 8dd6256
| Ok(Transformed::new(expr, false, TreeNodeRecursion::Jump)) | ||
| } | ||
| Expr::Alias(_) => Ok(Transformed::new( | ||
| expr.unalias(), |
|
Thanks again everyone for all the reviews |
…planning speed improvement) (apache#10835) * Stop copying LogicalPlan and Exprs in `CommonSubexprEliminate` * thread transformed * Update unary to report transformed correctly * Preserve through window transforms * track aggregate * Avoid re-computing Aggregate schema * Update datafusion/optimizer/src/common_subexpr_eliminate.rs * Avoid unecessary setting transform flat * Cleanup unaliasing
…planning speed improvement) (apache#10835) * Stop copying LogicalPlan and Exprs in `CommonSubexprEliminate` * thread transformed * Update unary to report transformed correctly * Preserve through window transforms * track aggregate * Avoid re-computing Aggregate schema * Update datafusion/optimizer/src/common_subexpr_eliminate.rs * Avoid unecessary setting transform flat * Cleanup unaliasing
Draft: this is still a draft as the performance doesn't really change -- I need to investigate more
Which issue does this PR close?
Closes #9873
Closes #9637 -- the last one of my planned PRs to make DataFusion planning faster by not copying so much
Note that we have more plans for CSE here: #10426
Rationale for this change
For low latency use cases the overhead of planning can be substantial, especially for relatively complex plans
What changes are included in this PR?
CommonSubexprEliminateto avoid deep copying plans using TreeNode APIAre these changes tested?
Functionally by existing CI
Performance tests: show a 2% improvement for
tpch_alland a 3% improvement fortpcds_allDetails
Are there any user-facing changes?
Performance Improvements (see above)