What methods does Column have?

[MarcoGorelli]

Currently, nothing is defined

dataframe-api/spec/API_specification/dataframe_api/column_object.py

Lines 1 to 2 in 9673b35

	class Column:
	pass

Just for my understanding, is this because it hasn't been done yet, or because it would be the same as from the Protocol?

https://github.com/data-apis/dataframe-api/blob/main/protocol/dataframe_protocol.py

[honno]

Just for my understanding, is this because it hasn't been done yet, or because it would be the same as from the Protocol?

https://github.com/data-apis/dataframe-api/blob/main/protocol/dataframe_protocol.py

Hasn't been done yet. Its worth noting the df interchange protocol specifies columns used in interchange (say df.__dataframe__().get_column(...)) and doesn't have to be supported by a libraries "top-level" column object (i.e. pd.Series).

[rgommers]

Hasn't been done yet.

Not quite. From gh-50, which is quite relevant here: We had (/have) a pretty strong consensus that there should not be a separate Series-like object, but only a DataFrame object with a single column. IIRC the key issue is that Series and DataFrame have so much API duplication, for little benefit. And that statement goes all the way back to gh-6, which contains a lot of the early discussion. The duplication pandas & co have with essentially duplicate APIs on the dataframe and series objects seemed undesirable to most folks.

That said, when actually implementing the interchange protocol we figured out that it's not 100% practical to not have a Column object at all. So the remaining question is what needs adding to a Column for an API, and where a single-column dataframe will do just fine.

[jorisvandenbossche]

While those older github issues indeed discussed to not have a separate 1D / column object, I seem to remember that on one of the more recent meetings where we discussed this, we landed on the opposite conclusion?

[jorisvandenbossche]

See https://github.com/data-apis/workgroup/blob/873d37ae66b1984fb3085cc1d5fe8634f3152aee/meeting_minutes/workgroup_meeting_2022_09_29.md?plain=1#L159-L277

[MarcoGorelli]

Is it enough to return a 1-column dataframe?

Say someone has the following (very common) pattern in pandas:

In [33]: df = pd.DataFrame({'a': [1,2,3], 'b': [4,5,6]})

In [34]: mask = df['a'] > 1

In [35]: df.loc[mask, ]
Out[35]:
   a  b
1  2  5
2  3  6

How could this be written with the standard? If Column supported __gt__, say, then the following would work:

df_standard = dataframe_standard(df)
mask = df.get_column_by_name('a') > 1
df_standard.get_rows_by_mask(mask)

Without a Column, or without Column.__gt__, what would the alternative be?

[rgommers]

While those older github issues indeed discussed to not have a separate 1D / column object, I seem to remember that on one of the more recent meetings where we discussed this, we landed on the opposite conclusion?

Right, indeed - thank you for finding that discussion. I think we need to update gh-50 with a clear summary of that. The tl;dr would be "so column is array-like and then has special behavior for missing values in reductions (with uniform skip_nulls kwarg across all functions), etc."

[rgommers]

I think we need to update gh-50 with a clear summary of that.

This is done now.

I believe this particular is resolved, since we have a good collection of methods on Column. So I'll close this issue.