[Minor] Short circuit ApplyFunctionRewrites if there are no function rewrites

[gruuya]

Which issue does this PR close?

Relates to #9373 and #9375.

Rationale for this change

I'm dealing with a situation where we have deeply nested plans, which we want to execute and stream the data into storage (Parquet/Delta), and we're hitting the stack overflow problem observed in the aforementioned issues.

Since this is on a write path we don't really need the analyzer/optimizer rules (I think), which are a part of the problem due to tree node recursion that takes place there. This is not an issue, since those can easily be opted out of via with_analyzer_rules/with_optimizer_rules.

However, the tightest bottleneck as per lldb is actually ApplyFunctionRewrites, which can't be opted out of, even though after #11155 it has no rewrite rules by default.

What changes are included in this PR?

Make ApplyFunctionRewrites simply bail out of the plan transformation/rewrite if it has no rules to apply (the default case presently).

Are these changes tested?

I wanted to add a test that checks for reference equity of the in/out plans but then recalled AnalyzerRule::analyze takes ownership of it.

So the only test is that I see a higher stack overflow threshold with this change.

Are there any user-facing changes?

None.

[jayzhan211]

datafusion/datafusion/optimizer/src/analyzer/mod.rs

Lines 139 to 141 in f044bc8

	let expr_to_function: Arc<dyn AnalyzerRule + Send + Sync> =
	Arc::new(ApplyFunctionRewrites::new(self.function_rewrites.clone()));
	let rules = std::iter::once(&expr_to_function).chain(self.rules.iter());

Is it better to opt it out here?

[gruuya]

datafusion/datafusion/optimizer/src/analyzer/mod.rs

Lines 139 to 141 in f044bc8

	let expr_to_function: Arc<dyn AnalyzerRule + Send + Sync> =
	Arc::new(ApplyFunctionRewrites::new(self.function_rewrites.clone()));
	let rules = std::iter::once(&expr_to_function).chain(self.rules.iter());

Is it better to opt it out here?

Makes sense to me, pushed the update.

[jayzhan211]

👍

[jayzhan211]

Thanks @gruuya

[alamb]

However, the tightest bottleneck as per lldb is actually ApplyFunctionRewrites, which can't be opted out of, even though after #11155 it has no rewrite rules by default.

Maybe we should simply remove this API -- people can write their own rules directly 🤔