waterWang opened a new pull request, #24435:
URL: https://github.com/apache/datafusion/pull/24435
## Which issue does this PR close?
Closes #24432.
## What changes are included in this PR?
This PR fixes the grouped floating-point MIN and MAX accumulators to use
`total_cmp` instead of `partial_cmp` for Float16, Float32, and Float64.
**Problem:** `partial_cmp` returns `None` for NaN comparisons, causing the
grouped update closure to unconditionally overwrite the current value. This
makes results order-dependent: `MAX([NaN, 1.0])` = `1.0` but `MAX([1.0,
NaN])`
= `NaN`.
**Fix:**
1. New `primitive_float_max_accumulator` and
`primitive_float_min_accumulator`
macros that use `total_cmp` (IEEE 754 total ordering) instead of
`partial_cmp`.
2. Initial values changed from `$NATIVE::MIN` / `$NATIVE::MAX` (finite
constants) to `$NATIVE::NEG_INFINITY` / `$NATIVE::INFINITY` (the true
extrema of the total ordering), so infinities and NaNs are handled
correctly.
## Are these changes tested?
Existing tests cover the non-grouped behavior. The grouped accumulators
now match the same total-ordering semantics.
## Are there any user-facing changes?
Grouped `MIN`/`MAX` on Float16, Float32, and Float64 columns will now
return deterministic results regardless of input order, batching, or
partitioning.
**Before:** `MAX([NaN, 1.0])` = `1.0` (order-dependent)
**After:** `MAX([NaN, 1.0])` = `NaN` (deterministic)
--
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]
