[BUG] window on unbounded preceeding and unbounded following can produce incorrect results.

[revans2]

Describe the bug
I tried to do a very basic window operation. The simplest I could think of.

spark.time(spark.range(5L).selectExpr("SUM(1) OVER()").show(truncate=false))

This produces nulls for all values on the GPU, but on the CPU it is all 5s. I thought this might be related to ordering, but it is not.

spark.time(spark.range(5L).selectExpr("SUM(1) OVER(ORDER BY id ROWS BETWEEN unbounded preceding and unbounded following)").collect())

does the same thing.

It gets even odder with COUNT, that returns values, that are wrong, but the cast to string is all empty strings...

[revans2]

This is a cudf bug

rapidsai/cudf#9672

I'll see if I can come up with a fix, but if we cannot get a fix soon we will have to fall back to the CPU for any row based window query where the values could overflow. (which might be all of them? or might be anything larger than 1/2 Int.MaxValue) I need to do some math to figure it out.

[revans2]

The CUDF fix was just merged. I'll put up some tests soon.