andygrove commented on code in PR #6563:
URL: https://github.com/apache/datafusion-comet/pull/6563#discussion_r4260696890
##########
spark/src/main/scala/org/apache/comet/serde/arrays.scala:
##########
@@ -540,39 +540,73 @@ object CometSlice extends CometExpressionSerde[Slice] {
private[comet] object ArraySetSupport {
val floatingPointReason: String =
- "Floating-point elements match Spark's signed-zero and NaN semantics
natively only on " +
- "Spark 4.2.0, whose optimizer normalizes the arguments (SPARK-54918)"
-
- // The native kernels match Spark only when the plan has already normalized
the arguments, and
- // only Spark 4.2.0 does that (SPARK-54918). Earlier releases keep flat
signed zeros apart.
- // From 4.0.5, 4.1.4 and 4.2.1, SPARK-59602 normalizes during evaluation
instead, which the
- // native kernels do not match for NaN payloads or nested zeros. A top-level
- // KnownFloatingPointNormalized marker cannot replace the version check:
Spark also normalizes
- // CreateArray, If, CaseWhen, and Coalesce recursively without wrapping the
resulting array.
- def normalizesArgumentsInPlan(version: String): Boolean =
- Utils.majorMinorPatchVersion(version).contains((4, 2, 0))
+ "Floating-point elements match Spark's signed-zero semantics natively only
on Spark " +
+ "4.0.5+, 4.1.4+ and 4.2+, which treat -0.0 and 0.0 as one value in these
functions " +
+ "(SPARK-54918, SPARK-59602)"
+
+ val collationReason: String =
+ "Elements that hold both a floating-point value and a non-UTF8_BINARY
collated string fall " +
+ "back to Spark, which compares the strings under their collation, while
Comet's native " +
+ "kernels compare their raw bytes"
+
+ // Spark 4.2.0 normalizes the arguments of these functions in the plan
(SPARK-54918), and 4.0.5,
+ // 4.1.4 and 4.2.1 normalize while evaluating them (SPARK-59602). Either
way, Spark treats -0.0
+ // and 0.0, and every NaN, as one value at any depth, which the spark_
variants match. Earlier
+ // releases keep -0.0 and 0.0 apart in a flat array. The check reads the
version rather than a
+ // KnownFloatingPointNormalized marker, because SPARK-59602 adds no marker,
and SPARK-54918
+ // normalizes CreateArray, If, CaseWhen and Coalesce without wrapping the
resulting array.
+ def normalizesFloats(version: String): Boolean =
+ Utils.majorMinorPatchVersion(version).exists {
+ case (4, 0, patch) => patch >= 5
+ case (4, 1, patch) => patch >= 4
Review Comment:
The `ARRAY<DOUBLE>` test is in 10e306072c, and `main` is merged now that
#6716 has landed. `array_null_short_circuit_float.sql` runs your `slice(b, 0,
1)` query over `array<double>` columns, with the start written as `IF(a IS
NULL, 0, 1)` so only the rows whose `a` is NULL have the invalid one, plus a
second query whose right operand is an ANSI cast of a malformed string. Both
expect `array_union` to run natively. The fixture opts `ArrayUnion` in, so it
exercises `spark_array_union` on every Spark version and not only where float
elements run natively by default.
With `withNullShortCircuit` removed from `CometArrayUnion`, the `slice`
query fails with `INVALID_PARAMETER_VALUE.START` and the cast query with
`CAST_INVALID_INPUT`. With it, both match Spark. I ran the fixture on 3.4, 3.5,
4.0, 4.1 and 4.2. The conflict in `CometArrayUnion.convert` kept this PR's
`ArraySetSupport.function` routing and #6716's `withNullShortCircuit`, the way
`CometArrayRemove` and `CometArrayContains` combine them for their `spark_`
variants.
That was the last item in your review, so I think this is ready for another
look.
--
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]
