[EPIC] Stop copying LogicalPlan during OptimizerPasses

@mustafasrepo

Is your feature request related to a problem or challenge?

Broken out from #9577 where @mustafasrepo @comphead and @jayzhan211 and I were discussing optimizer performance

TLDR is that the datafusion optimizer is slow. When I did some profiling locally by running the following

cargo bench --bench sql_planner -- physical_plan_tpch_all

My analysis is that almost 40% of the planning time is spent in SimplifyExprs and CommonSubexprEliminate and most of that time is related to copying expressions from what I can tell

While those passes themselves internally make a bunch of clones, which we are improving (e.g. @jayzhan211 on #9628) I think there is a more fundamental structural problem

I think a core challenge is that the OptimizerRule trait pretty much requires copying Exprs on each pass, as it gets a &LogicalPlan input, but produces a LogicalPlan output

    // Required methods
    fn try_optimize(
        &self,
        plan: &LogicalPlan,
        config: &dyn OptimizerConfig
    ) -> Result<Option<LogicalPlan>, DataFusionError>;

This mean any pass that works on Exprs must clone all Exprs (by calling LogicalPlan::expressions()) rewrite them, and then then create a new LogicalPlan with those new Exprs.

Here is that pattern in the expression simplifier:

https://github.com/apache/arrow-datafusion/blob/0eec5f8e1d0f55e48f5cdc628fbb5ddd89b91512/datafusion/optimizer/src/simplify_expressions/simplify_exprs.rs#L112-L123

Describe the solution you'd like

Find some way to avoid clone'ing exprs during LogicalPlan rewrite

Update: here are the tasks:

Infrastructure Preparation

Update OptimizerRules to avoid copying

Update AnalyzerRules to avoid copying

AnalyzerMisc: Minor: Avoid copying all expressions in Analzyer / check_plan #9974
InlineTableScan: Avoid copies in InlineTableScan via TreeNode API #10038
ApplyFunctionRewrites: fix NamedStructField should be rewritten in OperatorToFunction in subquery regression (change ApplyFunctionRewrites to use TreeNode API #10032
TypeCoercion: Stop copying LogicalPlan and Exprs in TypeCoercion #10210
TypeCoercion more: Onyl recompute schema in TypeCoercion when necessary #10365
CountWildcardRule: (needs a little reworking to avoid clones) Avoid copies in CountWildcardRule via TreeNode API #10066

Update Other to avoid copying

Stop copying plans in LogicalPlan::with_param_values #10016

Describe alternatives you've considered

No response

Additional context

We have talked about various other ways to reduce copying of LogicalPlans as well as its challenges in other tickets:

I have two idea before deep diving to the code

rewrite to have owned LogicalPlan then we can have owned expression easily. <- lots of code change
lazy clone on rewrite, the spirit similar to copy on write. The idea from simplify() Add a ScalarUDFImpl::simplfy() API, move SimplifyInfo et al to datafusion_expr #9304 may help here too or we can rely on Transformed<Expr>

enum SimpliedExpr {
  Simplified(Expr)
  Original
}

fn simply_expr(&self, expr: &Expr) -> SimpliedExpr

if expr is Original, we can avoid clone.

I have two idea before deep diving to the code

I think 2 sounds interesting

Another think I was thinking was something like LogicalPlan::rewrite(mut self, rewriter)

I think that along with Arc::try_unwrap could be used to minimize the places where cloning was actually needed

Maybe we can prototype with

impl OptimizerRule {
...

  /// does this rule support rewriting owned plans?
  fn supports_owned(&self) -> bool { return false }

  /// if supports_owned returns true, calls try_optimize_owned
  fn try_optimize_owned(
        &self,
        plan: LogicalPlan,
        config: &dyn OptimizerConfig
    ) -> Result<Transformed<LogicalPlan>, DataFusionError> {
      not_implemented_err!("try_optimized_owned is not implemented for this rule")
    }
...

And then play around with the code that calls try_optimize here https://github.com/apache/arrow-datafusion/blob/f4107d47bb4c0260d301294ddfc7c67d96574636/datafusion/optimizer/src/optimizer.rs#L360-L368 to try and use the try_optimize_owned API (without having to rewrite all the optimizers) for SimplifyExprs

If we could show a significant performance improvement for the sql_planner benchmarks then I think it would be worth spending time reworking the other APIs

@alamb

@alamb I think being able to skip failed rules is the main challenge to rewriting the plan with ownership

https://github.com/apache/arrow-datafusion/blob/37253e57beb25f0f1a4412b75421a489c2cb3c6a/datafusion/optimizer/src/optimizer.rs#L325

We need to restore the original plan if the rewrite fails at some point, however, the old plan is consumed and we lose ownership. Giving up ownership too early is not a good idea if we need to restore the original plan.

Instead of Transformed, I think we need to preserve the old data and fail state to mimic Err in Result but with old data

#[derive(Debug, PartialEq, Eq)]
pub enum OptimizedState {
    Yes,
    No,
    Fail,
}

#[derive(Debug)]
pub struct Optimized<T, E = DataFusionError> {
    pub optimzied_data: Option<T>,
    // Used to store the original data if optimized successfully
    pub original_data: T,
    pub optimized_state: OptimizedState,
    // Used to store the error if optimized failed, so we can early return but preserve the original data
    pub error: Option<E>,
}

impl<T, E> Optimized<T, E> {
    pub fn yes(optimzied_data: T, original_data: T) -> Self {
        Self {
            optimzied_data: Some(optimzied_data),
            original_data,
            optimized_state: OptimizedState::Yes,
            error: None,
        }
    }

    pub fn no(original_data: T) -> Self {
        Self {
            optimzied_data: None,
            original_data,
            optimized_state: OptimizedState::No,
            error: None,
        }
    }

    pub fn fail(original_data: T, e: E) -> Self {
        Self {
            optimzied_data: None,
            original_data,
            optimized_state: OptimizedState::Fail,
            error: Some(e),
        }
    }
}

Then, for every optimization, we return Optimized<LogicalPlan> or Optimzied<Expr>

@alamb

@alamb I think being able to skip failed rules is the main challenge to rewriting the plan with ownership

This is an excellent point

Instead of Transformed, I think we need to preserve the old data and fail state to mimic Err in Result but with old data

If we need to preserve the old data, there is no choice but to copy the LogicalPlan on each pass 🤔 (well I guess there could be some way to repair the partially rewritten plan, but that sounds very compliated).

Since skip_failed_rules is false by default:

https://github.com/apache/arrow-datafusion/blob/dcfe70987e98da0410146b5e1292ab20f3f118e0/datafusion/common/src/config.rs#L549

Could we maybe only do the copy when skip_failed_rules is enabled? Something like this:

if skip_failed_rules {
  let original_plan = plan.clone();
  let new_plan = optimizer.try_optimize_owned(plan)
   .ok_or_else(|e| original_plan);
} else {
  optimizer.try_optimize_owned(plan)?
}

I think it is possible to preserve unnecessary clones even if we preserve the old plan, and only clone the data for the new expr and new plan, but I agree it is a little complicated, it needs the enum I show above. We can optimize for the default path first, and others later on.

Let's give it a try!

note: I profile my change and find out the time does not improve at all, I find that the cloned inside expr_list_to_fields may be the core issue.

I analyze the sql_planner again and find that exprlist_to_fields and calc_func_dependencies_for_project are the two most time spending func.

so I think #9595 might be an important step for the optimization here.

https://github.com/apache/arrow-datafusion/blob/ad8d552b9f150c3c066b0764e84f72b667a649ff/datafusion/expr/src/logical_plan/plan.rs#L1804-L1813

I played around with some ideas last night and it is looking promising (though not yet done). I put my draft here #9708. I hope to try and work on it more over the next few days, but I am pretty busy preparing for meetups / presentations / papers for next week. It may be a few more days

I would be surprised if #9708 improves, let me see if I can improve #9658 based on it.

BTW here is a video of how I profile DataFusion: #9577 (comment)

I'm curious, It seems like most of the reasoning behind Arc<LogicalPlan> and clone are due to the optimizer stage. Polars uses an intermediate representation ALogicalPlan (arena allocated logical plan) solely for the optimizer stage. I wonder if the same techinque could be applied here.

I think area allocating is pretty common in various compiler/query optimizer systems because you know the lifetime of a plan doesn't outlive the the optimizer pass (as in you can bound the intermediate results)

I think rust actually is ideally setup to avoid having to arena allocate (which still requires copying stuff to a new plan) (using ownership) -- we just need to regigger how DataFusion is setup to use this. I think we are close

There is one issue that we are not able to call Arc::into_inner for skip-failed-rules path since we hold a clone for restoration.

@alamb

@alamb, @jayzhan211 this is another low hanging fruit to refactor UnwrapCastInComparison: #10087

I am interested in refactoring some of the OptimizeRule, I'll start with some easy optimization rules

@Lordworms

Thanks @Lordworms -- that sounds great. There are now some good examples of the basic pattern (e.g #10081 and #10066)

@peter-toth

Update here is that I am hacking on getting TypeCoercion (#10039) to avoid clones. It is turning out to be tricker than others as type coercion actually updates the types of the plans, so when rewriting expressions I need to find some way to re-calculate the schemas correctly. I am continuing to hack on that as @peter-toth works on CSE (#10067) and @Lordworms is cranking on some of the other passes

I also re-profiled the tpcds benchmarks via

cargo bench --bench sql_planner -- tpcds_all

And here is what it shows:

Thus I am going to take a peek to see if there is any low hanging fruit in ApplyFunctionRewrites EliminateCrossJoin, OptimzieProjections and FilterPushdown

I filed some tickets to track the larger items #10209 #10210 #10211

take eliminate_duplicated_expr and eliminate_cross_join

@Lordworms

take eliminate_duplicated_expr and eliminate_cross_join

FYI @Lordworms I took a look at brief look and they may be somewhat tricky. Of the two, i think elimnate_duplicated_expr is the easier

@Lordworms

take eliminate_duplicated_expr and eliminate_cross_join

FYI @Lordworms I took a look at brief look and they may be somewhat tricky. Of the two, i think elimnate_duplicated_expr is the easier

Yes, actually I have finished the elimnate_duplicated_expr already, I'll merge them two as a single PR after finishing eliminate_cross_join

If possible two single PRs would be easier to review and merge

If possible two single PRs would be easier to review and merge

sure

I filed tickets for the other optimizer passes and linked them to the description of this ticket

@appletreeisyellow

Update here is that this issue is done other other than

Stop copying LogicalPlan and Exprs in SingleDistinctToGroupBy #10295 (which @appletreeisyellow is working on )
Stop copying Exprs and LogicalPlans so much during Common Subexpression Elimination #9873 which I will do (but I am currently waiting for @peter-toth 's work in Rewrite CommonSubexprEliminate to avoid copies using TreeNode #10067 to complete)

Navigation Menu

Repository navigation

[EPIC] Stop copying `LogicalPlan` during OptimizerPasses #9637

Is your feature request related to a problem or challenge?

Describe the solution you'd like

Describe alternatives you've considered

Additional context

21 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[EPIC] Stop copying LogicalPlan during OptimizerPasses #9637

Description

Is your feature request related to a problem or challenge?

Describe the solution you'd like

Describe alternatives you've considered

Additional context

Activity

jayzhan211 commented on Mar 16, 2024

alamb commented on Mar 16, 2024

jayzhan211 commented on Mar 17, 2024

alamb commented on Mar 17, 2024

jayzhan211 commented on Mar 17, 2024

alamb commented on Mar 17, 2024

jayzhan211 commented on Mar 17, 2024

jayzhan211 commented on Mar 20, 2024

alamb commented on Mar 20, 2024

jayzhan211 commented on Mar 20, 2024

alamb commented on Mar 20, 2024

universalmind303 commented on Mar 20, 2024

alamb commented on Mar 20, 2024

jayzhan211 commented on Mar 21, 2024

21 remaining items

peter-toth commented on Apr 15, 2024

Lordworms commented on Apr 17, 2024

alamb commented on Apr 19, 2024

alamb commented on Apr 23, 2024

alamb commented on Apr 23, 2024

alamb commented on Apr 24, 2024

Lordworms commented on Apr 24, 2024

alamb commented on Apr 24, 2024

Lordworms commented on Apr 24, 2024

alamb commented on Apr 24, 2024

Lordworms commented on Apr 24, 2024

alamb commented on Apr 29, 2024

alamb commented on May 17, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

[EPIC] Stop copying `LogicalPlan` during OptimizerPasses #9637