Bugfix: Calling count with None arguments

[timsaucer]

Which issue does this PR close?

None

Rationale for this change

This is a two line bugfix.

What changes are included in this PR?

If the user calls count() it should internally use count_star()

Are there any user-facing changes?

None

[Michael-J-Ward]

In #771, I flirted with the idea of removing count_star from the Rust side, but then added it back in for this release. It's really just porcelain for count(lit(1)) (I think).

Now that we have the benefit of the python wrappers, do you think we should try to push aliases such as this into the python side?

EDIT:

Thinking a little more, would count(None, distinct = True) be a potential foot-gun that we should avoid, anyway?

The other thing coming out of #771 is that we could use a better AggregateBuilder API / wrapper around the new AggregateExt trait

[timsaucer]

In regards to aliases, I do think putting them on the wrappers makes sense. One thing I didn't do with #750 was to remove the alias function definitions on the rust side, even when there is an alias in the python wrapper. I'll add an issue to make sure that gets cleaned up.

Yes, count(None, distinct = True) would be a problem. What do you think - should we update count_star or maybe just add to the docstring to say that None arguments is a special case that ignores distinct?

In regards to AggregateBuilder would you be willing to wait one more release? We just merged in apache/datafusion#11550 which I'm guessing will be in datafusion 41.0. With that we get both the aggregate functions and window functions using the same API. That might make the minimal amount of churn. But if you do want to get in the aggregates for datafusion-python 40.0 I can probably knock that out in the next couple of days.

[Michael-J-Ward]

In regards to AggregateBuilder would you be willing to wait one more release? We just merged in apache/datafusion#11550 which I'm guessing will be in datafusion 41.0.

First, excellent upstream PR. I literally have "improve upstream Builder API" in my post upgrade hit-list, so I'm very happy to see that you've knocked it out.

Second, we can absolutely wait till 41. I've been slow to the 40 upgrade, so let's get it out the door first.

What do you think - should we update count_star or maybe just add to the docstring to say that None arguments is a special case that ignores distinct?

Could count(None, distinct = True) map to count(lit(1), distinct=True) instead of count_star()? And before that, do we need to support count(None) at all?

[timsaucer]

Excellent suggestion. Pushed the change.