Add set based operations for arrays: array_intersect, array_union, array_except, and arrays_overlap for running on GPU

[NVnavkumar]

Fixes #4932, #5228 , #5188, and #5222. Requires rapidsai/cudf#11043 and rapidsai/cudf#11143 to be merged.

This implements the SQL functions for set-based operations on arrays array_intersect, array_union, array_except, and arrays_overlap for running on the GPU. A few caveats:

1. There is a minor bug ([BUG] Exception calling collect() when partitioning using with arrays with null values using array_union(...) #5957) that came up when testing array_union, and it looks like a bug in the partitioning code when using collect()

2. The 3 operations array_intersect, array_union and array_except cannot guarantee the same order when used on the GPU vs the CPU (because they are set-based operations), so we wrap these operations in sort_array(...) call for testing purposes.

3. arrays_overlap returns a boolean result value, so there is no need to sort, but the implementation is complicated because of these requirements for the function implementation (see https://spark.apache.org/docs/3.2.1/api/sql/index.html#arrays_overlap)

arrays_overlap(a1, a2) - Returns true if a1 contains at least a non-null element present also in a2. If the arrays have no common element and they are both non-empty and either of them contains a null element null is returned, false otherwise.

[ttnghia]

There should be an empty line at the end of the file as an implicit convention.

Suggested change

	)
	)

[NVnavkumar]

Another example:

scala> val df = List((Array(Double.NaN), Array(-Double.NaN))).toDF("a","b")
df: org.apache.spark.sql.DataFrame = [a: array<double>, b: array<double>]

scala> df.selectExpr("array_union(a, b)").collect
res1: Array[org.apache.spark.sql.Row] = Array([WrappedArray(NaN)])

[ttnghia]

What's wrong with my spark-shell (spark 3.1.2)?

scala> spark.sql("SELECT array_union(array(cast('nan' as double)), array(-cast('-nan' as double)))").show
+------------------------------------------------------------------------+
|array_union(array(CAST(nan AS DOUBLE)), array((- CAST(-nan AS DOUBLE))))|
+------------------------------------------------------------------------+
|                                                             [NaN, null]|
+------------------------------------------------------------------------+


scala> spark.sql("SELECT array_union(array(cast('nan' as double)), array(-cast('nan' as double)))").show
+-----------------------------------------------------------------------+
|array_union(array(CAST(nan AS DOUBLE)), array((- CAST(nan AS DOUBLE))))|
+-----------------------------------------------------------------------+
|                                                             [NaN, NaN]|
+-----------------------------------------------------------------------+


scala> val df = List((Array(Double.NaN), Array(-Double.NaN))).toDF("a","b")
df: org.apache.spark.sql.DataFrame = [a: array<double>, b: array<double>]

scala> df.selectExpr("array_union(a, b)").collect
res17: Array[org.apache.spark.sql.Row] = Array([WrappedArray(NaN, NaN)])

[ttnghia]

@NVnavkumar Did you generate your last example by running on the GPU? I did mine on the CPU. If the correct behavior for comparing NaNs is that they are always considered unequal then I need to fix the JNI layer to pass in the correct NaN comparison parameter.

[NVnavkumar]

@NVnavkumar Did you generate your last example by running on the GPU? I did mine on the CPU. If the correct behavior for comparing NaNs is that they are always considered unequal then I need to fix the JNI layer to pass in the correct NaN comparison parameter.

Both examples were on the CPU, Spark 3.2.1

[NVnavkumar]

Tried with Spark 3.1.3

scala> val df = List((Array(Double.NaN), Array(-Double.NaN))).toDF("a","b")
df: org.apache.spark.sql.DataFrame = [a: array<double>, b: array<double>]

scala> df.selectExpr("array_union(a, b)").collect
res0: Array[org.apache.spark.sql.Row] = Array([WrappedArray(NaN)])

But in Spark 3.1.1, there is this behavior:

scala> val df = List((Array(Double.NaN), Array(-Double.NaN))).toDF("a","b")
df: org.apache.spark.sql.DataFrame = [a: array<double>, b: array<double>]

scala> df.selectExpr("array_union(a, b)").collect
res0: Array[org.apache.spark.sql.Row] = Array([WrappedArray(NaN, NaN)])

[ttnghia]

Do we have any config for this? If we can retrieve the config in Spark then we can select the right parameter to call cudf API.

Similar to this: https://github.com/NVIDIA/spark-rapids/blob/f2d6157251dae10874e868337213c150d080048c/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuOverrides.scala#L3368

[NVnavkumar]

This is interesting, because in the JDK itself, Double.doubleToLongBits actually shortcuts the NaN condition:

https://github.com/openjdk/jdk/blob/a0a0539b0d3f9b6809c9759e697bfafd7b138ec1/src/java.base/share/classes/java/lang/Double.java#L860-L865

So I guess Spark wasn't using it before 3.1.3

[NVnavkumar]

Do we have any config for this? If we can retrieve the config in Spark then we can select the right parameter to call cudf API.

Similar to this:

https://github.com/NVIDIA/spark-rapids/blob/f2d6157251dae10874e868337213c150d080048c/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuOverrides.scala#L3368

It looks like they switched to SQLOpenHashSet in 3.1.3 and added the NaN check:

https://github.com/apache/spark/blob/d1f8a503a26bcfb4e466d9accc5fa241a7933667/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/collectionOperations.scala#L3215

Here's the PR: apache/spark#33955 and the original Spark issue with NaN: https://issues.apache.org/jira/browse/SPARK-36702

[NVnavkumar]

I think I will try to xfail the test for 3.1.1 and 3.1.2 and file a follow up issue for the NaN case.

[NVnavkumar]

build

[ttnghia]

build

[revans2]

My main concern is documentation. We have some new functionality that is not 100% compatible with Spark. We should at a minimum document that it is not compatible and exactly in what ways it is not compatible. Ideally if there are issues with how Spark does something, like with -0.0 and we have filed an issue for it with Spark we should include all of that in the documentation.

[NVnavkumar]

build