REF: check instead of catching TypeError

[jbrockmendel]

No description provided.

[jorisvandenbossche]

What's an example that now ran into this TypeError?

[jbrockmendel]

What's an example that now ran into this TypeError?

results = [3, 2, 1.5]; keys = ['size', 'count', 'mean']

[jorisvandenbossche]

Sorry, I meant the original dataframe / groupby example or test case that gives those values at this location

[jbrockmendel]

Identifying some relevant cases with an assert False, the place to look is in tests.frame.test_apply, e.g. test_non_callable_aggregates

[jreback]

thanks @jbrockmendel

[jorisvandenbossche]

@jreback if I review a PR with a comment, can you try to give me a little bit of time to respond to Brock's response?

@jbrockmendel thanks for that reference. So in that test case, you have functions that return single numbers.
But, you can also have an aggregation that returns an array-like or list, which has a dimension, but cannot be concatted. Quick dummy example with the case from the tests:

In [8]: df = DataFrame( 
   ...:     {"A": [None, 2, 3], "B": [1.0, np.nan, 3.0], "C": ["foo", None, "bar"]} 
   ...: )  

In [9]: df.agg([lambda x: list(x.unique())]) 

Out[9]: 
                        A                B                 C
<lambda>  [nan, 2.0, 3.0]  [1.0, nan, 3.0]  [foo, None, bar]

This now fails.
(the above example looks a bit unnatural, but in general, returning lists from an aggregation is something that certainly gets done I think, also in groupby context)

[jreback]

@jorisvandenbossche i gave a day here which is reasonable
if you want to block then put a hold on the review

[jbrockmendel]

@jreback while i agree an explicit "hold" would be helpful, I think Joris gets some extra slack because of the timezone differences.

@jorisvandenbossche have i mentioned recently how much I dislike listlike corner cases? But the case you brought is Technically Correct, so I'd be OK if you want to revert (and add that test case)

[jorisvandenbossche]

@jreback while i agree an explicit "hold" would be helpful, I think Joris gets some extra slack because of the timezone differences.

Indeed ;) For me, the "1 day" (15 hours since Brocks comment) was just a morning after having slept.
Maybe just interpret a comment that is not accompanied with a github approve as a sign that I want to check back?

But the case you brought is Technically Correct, so I'd be OK if you want to revert (and add that test case)

Do you directly think of another way to do this that you would prefer more than the try/except? Otherwise I can indeed do a revert yes.

[jbrockmendel]

Do you directly think of another way to do this that you would prefer more than the try/except?

The check in this PR was the best idea I had. Do you have any suggestions?

[jorisvandenbossche]

We could check for actual series/frame objects (since concat only allows that), but personally, I find this a perfect use case for the try except.

[jorisvandenbossche]

Added a test case here: #29632