asugranyes opened a new pull request, #53695:
URL: https://github.com/apache/spark/pull/53695
### What changes were proposed in this pull request?
Add normalization of -0.0 to 0.0 in hash-based array operations:
`array_distinct`, `array_union`, `array_intersect`, and `array_except`.
Changes:
- Add `normalizeZero()` and `normalizeZeroCode()` to `SQLOpenHashSet` for
interpreted and codegen paths
- Apply normalization in all four array operations before hashing
### Why are the changes needed?
IEEE 754 defines -0.0 == 0.0, but they have different binary representations
and hash codes. This causes incorrect behavior when arrays contain both values:
```scala
// Returns [0.0, -0.0, 1.0] instead of [0.0, 1.0]
Seq(Array(0.0, -0.0,
1.0)).toDF("values").selectExpr("array_distinct(values)").show()
// Returns [0.0, -0.0] instead of [0.0]
Seq((Array(0.0), Array(-0.0))).toDF("a", "b").selectExpr("array_union(a,
b)").show()
// Returns [] instead of [0.0]
Seq((Array(0.0, 1.0), Array(-0.0, 2.0))).toDF("a",
"b").selectExpr("array_intersect(a, b)").show()
// Returns [0.0, 1.0] instead of [1.0]
Seq((Array(0.0, 1.0), Array(-0.0))).toDF("a",
"b").selectExpr("array_except(a, b)").show()
```
Spark already normalizes -0.0 to 0.0 in join keys, window partition keys,
and aggregate grouping keys via `NormalizeFloatingNumbers`. This fix makes
array operations consistent.
### Does this PR introduce _any_ user-facing change?
Yes. Array operations now correctly treat -0.0 and 0.0 as equal, consistent
with SQL semantics and IEEE 754.
### How was this patch tested?
- Added unit tests in `CollectionExpressionsSuite` for all four operations
with Double and Float types
- All existing tests pass
### Was this patch authored or co-authored using generative AI tooling?
No
--
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]
