feat: Add eliminate group by constant optimizer rule

[korowa]

Which issue does this PR close?

Closes #.

Rationale for this change

Initial intention was to improve clickbench q34 -- it contains aggregation by constant and URL which makes aggregation switch to row-based mode which is slower that single field group by.

In general, elimination of constants from GROUP BY seems to be reasonable optimization.

What changes are included in this PR?

New logical optimizer rule which removes constant or originated from constant expressions from Aggregate::group_expr, and places additional projection on top of aggregation, to satisfy original schema for downstream operators.

Optimizer rule is placed between last SimplifyExpressions and OptimizeProjections, since it requires constant evaluation, and projections, produced by this rule, may be merged.

Are these changes tested?

New unit tests and sqllogictests

Are there any user-facing changes?

No

[alamb]

This is a really nice contribution @korowa (as always). I think the code is well tested, documented and explained.

I had some suggestions, but I don't think they are required for merging.

Thanks again

[alamb]

this certainly looks like a plan improvement (fewer grouping keys) -- among other thing that would also allow this query to use the faster specialized group accumulator

[alamb]

I think you could use TreeNode::Exists here:

datafusion/datafusion/common/src/tree_node.rs

Line 408 in 045e8fc

fn exists<F: FnMut(&Self) -> Result<bool>>(&self, mut f: F) -> Result<bool> {

Which handles all types of expressions and would be general

Something like this, perhaps (untested)

fn is_constant_expression(expr: &Expr) -> bool {
  !expr.exsts(|expr| {
    match expr { 
      Expr::Column(_) => true,
      Expr::ScalarFunction(e) if matches!(
                e.func.signature().volatility,
                Volatility::Volatile
      ) => true,
      _ => false
 })

[korowa]

Thank you for the hint. I've checked id out, and seems like there is no TreeNode API method for this case, since it requires all arguments of scalar function to be literals, and exists will return on first match.

[alamb]

Since this PR looks good and makes improvements I am going to merge it in. @korowa perhaps you can consider the feedback on #10591 (comment) as a follow on PR

🚀