CuteChuanChuan opened a new pull request, #19994:
URL: https://github.com/apache/datafusion/pull/19994
## Which issue does this PR close?
- Closes #11570.
## Rationale for this change
The `CaseExpr` implementation is expensive. A common usage pattern
(particularly in TPC-DS benchmarks) is to protect against divide-by-zero:
```sql
CASE WHEN y > 0 THEN x / y ELSE NULL END
```
This entire expression can be replaced with a simpler divide operation that
returns NULL when the divisor is zero, avoiding the overhead of full CASE
evaluation.
## What changes are included in this PR?
1. New EvalMethod::DivideByZeroProtection variant - A specialization for the
divide-by-zero protection pattern
2. Pattern detection - Detects patterns like:
- CASE WHEN y > 0 THEN x / y ELSE NULL END
- CASE WHEN y != 0 THEN x / y ELSE NULL END
- CASE WHEN 0 < y THEN x / y ELSE NULL END
3. Critical validation - Ensures the divisor in the division matches the
operand being checked (addresses feedback from PR#12049)
4. Safe division implementation - Uses Arrow kernels to perform division
that returns NULL on zero:
- eq to create zero mask
- zip to replace zeros with ones (avoid division error)
- div to perform division
- nullif to set NULL where divisor was zero
## Are these changes tested?
Yes, added two new tests:
- test_divide_by_zero_protection_specialization - Verifies pattern is
detected and results are correct
- test_divide_by_zero_protection_specialization_not_applied - Verifies
optimization is NOT applied when divisor doesn't match checked operand (key
feedback from PR#12049)
## Are there any user-facing changes?
No. This is an internal optimization that produces the same results but with
better performance.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
