Discuss: transformation vs. aggregation in `agg` vs. `transform`

There have been multiple issues regarding

• Confusion about relative respective roles of agg and transform (and apply).
  Why do I get strange aggregation result from DataFrame groupby()? #26960, Behavior of new df.agg, df.transform and df.apply is very inconsistent #18103 in both the DataFrame and Grou
• transform('rank') and others returning the wrong answer (first issue from 2016): bug when filling missing values with transform? #14274, Shortcut functions in transform are not grouped #19354, Wrong output of GroupBy transform with string input (e.g., transform('rank')) #22509,

I'd like to prepare a PR to fix this, but I need to know what the consensus is first.

1. should Groupby/DataFrame/Series.agg disallow transformations?

#14741 (comment) about Groupby.agg

you need to use .transform. .agg is by definition a reducer.

agg currently accepts transformations as well.

DataFrame.agg was merged in #27389 despite repeated objections to mixing transformations and aggregations, #14668 (comment) and #14668 (comment).

2. What should transform('rank') return?

updated

g.min               # one value per group. ok.
g.transform('min')  # one value per group, broadcasted. ok.
g.rank()            # like-index result. Ok
g.transform('rank') # Currently, a nonsense result.

On the one hand, users have been told that transformations don't belong in transform (!): #22509 (comment), #14274 (comment). This makes sense if you think of the transform('name') form as solely for broadcasting aggregations.

On the other hand, the documentation for transform, as well as Wes Mckinney's excellent pandas book portray transform as the dedicated tool for shape-preserving operations, so excluding them from the transform('name') case would be a little surprising.

Personally, I'm +0 for deprecating transform('rank') and with a warning to use g.rank(), as well as for the other transformation ops.

[WillAyd]

I would be fine with the first point to ensure agg reduces. I'm not clear on what issue points 2 and 3 are trying to solve

[TomAugspurger]

I would be fine with the first point to ensure agg reduces

Agreed. We should deprecate the current behavior with a warning.

I also don't understand point 2.

3. Is broadcasting a transformation result useful in some other cases?

Yes, I find this useful.

I'm not clear on what issue points 2 and 3 are trying to solve

I also don't understand point 2.

Edited OP for clarity. What should g.transform('rank') return? should it be the same as g.rank() or be deprecated with a warning? currently it returns nonsense.

Agreed. We should deprecate the current behavior with a warning.

Good. But how does that sit with 25.0 being released and the no-deprecation policy during this release cycle coming up on 1.0? Can we do it?

[TomAugspurger]

Discussion on the mailing list: https://mail.python.org/pipermail/pandas-dev/2019-July/001030.html

I think the plan was to be light on new deprecations for 1.0. But I don't know how wedded we are to that idea.

[mroeschke]

I think the more recent discussion of this has moved to #35725, so closing in favor of that issue.