Potential optimization for CASE WHEN for protecting against divide by zero

[andygrove]

Is your feature request related to a problem or challenge?

A common usage of CASE WHEN (particularly in the TPC-DS benchmark) is to protect against divide by zero errors. For example:

CASE WHEN y > 0 THEN x / y ELSE null END

The CaseExpr implementation is quite expensive and we could replace this whole expression with a divide kernel that returns null if the right hand side is zero (Rust has a div_checked function that already provides this functionality).

arrow-rs already has the following code in arrow-array/src/arithmetic.rs, which is really close to what we need. I think we just need a version that returns an Option::None instead of an Err.

#[inline]
fn div_checked(self, rhs: Self) -> Result<Self, ArrowError> {
    if rhs.is_zero() {
        Err(ArrowError::DivideByZero)
    } else {
        self.checked_div(rhs).ok_or_else(|| {
            ArrowError::ComputeError(format!(
                "Overflow happened on: {:?} / {:?}",
                self, rhs
            ))
        })
    }
}

Describe the solution you'd like

No response

Describe alternatives you've considered

No response

Additional context

No response

[irenjj]

take

[irenjj]

I have tried to solve this problem, but encountered some issues: Should we simplify the when_expr and then_expr in case_when_no_expr to div_checked? If it should be handled in this function, how can we recognize it as div_checked?

[andygrove]

@irenjj I think it would be a case of writing code like this to detect this particular pattern:

            } else if when_then_expr.len() == 1 {
                let when_then = &when_then_expr[0];
                if let Some(b) = when_then.1.as_any().downcast_ref::<BinaryExpr>() {
                    if b.op() == Operator::Gt {
                        if let Some(lit) = when_then.1.as_any().downcast_ref::<Literal>() {
                            if matches!(lit.value(), ScalarValue::Int32(Some(0))) {
                                
                            }
                        }
                    }
                }

[CuteChuanChuan]

Hi @Jefffrey , could I try this?

[Jefffrey]

Hi @Jefffrey , could I try this?

Feel free to tackle this