[FEA] customizable default for segmented_reduce value on empty segments

[gerashegalov]

Is your feature request related to a problem? Please describe.
Currently segmented_reduce returns NULL for various inputs, including empty inputs. For empty inputs NULL is not always a natural choice. E.g. when implementing a Boolean aggregation exists NVIDIA/spark-rapids#4973 for Spark SQL, empty input quite intuitively results in false

Describe the solution you'd like
The most generic solution is to simply to allow the user to pass the initial value for reduction similar to Scala fold

Describe alternatives you've considered
As a workaround, we have to compute a Bool column for the post processing step of fixing up "wrong" results.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[jrhemstad]

Makes sense to me. No different than cub's segmented reduce taking an initial value.

[bdice]

Providing an initial value for the reduction is different than post-processing to replace null values. Which do you want? There's an implicit 'operator identity' right now as the default value, and its behavior wrt null inclusion/exclusion has the potential to change if a default value besides the identity is allowed.

For example, consider the operator +, initial value of 0 (the operator identity), and segment data [null]. Currently, the behavior is to return null for both null_policy::INCLUDE and null_policy::EXCLUDE. It seems that this proposed change could mean either of:

1. reduce(+, 0, [null], include) == null, reduce(+, 0, [null], exclude) == 0
   • Equivalent to left-fold with null semantics from the initial value
   • Note the behavior change from null to 0 for exclude!
2. reduce(+, 0, [null], include) == 0, reduce(+, 0, [null], exclude) == 0
   • Equivalent to trivial post-processing to replace null with 0

If it's (1) you're seeking, this could be in libcudf. If it's (2) you're seeking, I would recommend a trivial postprocessing.

Here's a larger matrix of examples (edit: I'm re-reading/editing this to ensure it's aligned with my expectations):

Click to expand

Initial value	Segment Data	Nulls	Current behavior	Result of left fold (+) from initial value	Result of post-processed null replacement
(identity)	[]	Include	null	0 (currently gives null!)	0
(identity)	[]	Exclude	null	0 (currently gives null!)	0
(identity)	[1, 2]	Include	3	3	3
(identity)	[1, 2]	Exclude	3	3	3
(identity)	[null]	Include	null	null	0
(identity)	[null]	Exclude	null	0 (currently gives null!)	0
(identity)	[1, null]	Include	null	null	0
(identity)	[1, null]	Exclude	1	1	1
42	[]	Include		42	42
42	[]	Exclude		42	42
42	[1, 2]	Include		45	???
42	[1, 2]	Exclude		45	???
42	[null]	Include		null	42
42	[null]	Exclude		42	42
42	[1, null]	Include		null	42
42	[1, null]	Exclude		43	???

[jrhemstad]

Providing an initial value for the reduction is different than post-processing to replace null values.

Unless I'm misunderstanding, I don't think @gerashegalov is asking for post-processing to replace nulls.

He's saying that when there is an empty segment it currently returns null. It would be nice instead if an empty segment would just return the init value instead.

This is a natural extension of howstd::reduce on an empty input range will just return the initial value.

https://godbolt.org/z/e8TYo4Yd6

I was a little surprised that we don't have an initial_value argument in cudf::reduce.

[gerashegalov]

Unless I'm misunderstanding, I don't think @gerashegalov is asking for post-processing to replace nulls.

Correct, we have already implemented postprocessing because there is no way to specify the initial value for aggregation https://github.com/NVIDIA/spark-rapids/blob/branch-22.06/sql-plugin/src/main/scala/com/nvidia/spark/rapids/higherOrderFunctions.scala#L398-L402

[davidwendt]

Both cudf::reduce and cudf::segmented_reduce accept an initial-value parameter now.
If you pass false as an initial value for the ANY aggregation type (cudf::make_any_aggregation<cudf::segmented_reduce_aggregation>()) then empty segments will return false and not null. Null rows are still controlled by the null_policy parameter.

input = [0,0], [1,1], [1,0], [0,1,1], [], [1,null], [0,null]
init_value = false
result = cudf::segmented_reduce(input, ANY, BOOL8, EXCLUDE, init_value)
result -> [false, true, true, true, false, true, false]

result = cudf::segmented_reduce(input, ANY, BOOL8, INCLUDE, init_value)
result -> [false, true, true, true, false, null, null]

Without the initial value, the empty segments will result in a null row.

[vyasr]

Can we close this now, then? Or is there an outstanding ask remaining?

[gerashegalov]

Can we close this now, then? Or is there an outstanding ask remaining?

Thanks, I don't believe so. We'll open more issues if needed