[BUG] segmented_reduce does not accept empty offsets

[ttnghia]

When the input is all empty, i.e., we have empty input values and empty offsets, segmented_reduce throws an exception:

C++ exception with description "CUDF failure at: ../../src/reductions/segmented/reductions.cpp:163: `offsets` should have at least 1 element."

This is wrong, as it should be able to handle the case of empty input.

[davidwendt]

This appears to be working as designed. There are a couple of tests for the case where in the input is empty. Here is one:

cudf/cpp/tests/reductions/segmented_reduction_tests.cpp

Line 1083 in d93e9c2

TEST_F(SegmentedReductionTestUntyped, EmptyInputWithOffsets)

I think it reasonable to return an exception when no offsets are provided.

[ttnghia]

I think that test had to generate offsets because of this bug, not because the API was designed so. If a column is empty (data), having empty offsets should be the correct behavior. We implemented make_empty_lists_column (and also empty strings column) by that design.

[davidwendt]

Nope, this is not a bug but was intentional behavior as per a request from Spark found here: #10556 (comment)

However, I think it reasonable to return an empty column if both the input and offsets are empty.
I can create a PR for that shortly.