Idk how to do this better:

• only simplify A.obs[k] to string? If so, should we allow to mix strings and AdRefs in by=[…]?
• simplify unless there are duplicate .idxs?
• don’t stringify at all and use AdRefs as df.columns?

Is the question here about the isinstance check? Like, what to do about people who pass in the index as a by column?

This line of code seems inoffensive other than I am not sure why ref.idx would be anything other than a str

the question is about “we now accept AdRefs, what to use as dataframe columns when they’re used”

I am not sure why ref.idx would be anything other than a str

>>> type(ad.acc.A.obsm["y"][:, 0].idx)
int

the question is about “we now accept AdRefs, what to use as dataframe columns when they’re used”

But these columns should really just be ref.idx since ad.acc.A.obsm["y"][:, 0] isn't a valid by unless I'm missing something.

why would it not be?

import numpy as np
import scanpy as sc
from anndata import acc, AnnData

adata = AnnData(np.random.rand(3, 4), obsm=dict(thing=np.array(["x", "y", "y"])))
sc.get.aggregate(adata, acc.A.obsm["thing"][:, 0], ())
# AnnData object with n_obs × n_vars = 2 × 4
#    obs: "A.obsm['thing'][:, 0]", 'n_obs_aggregated'

To clarify, if you leave it as-is, I think that's fine. If you revert to the old behavior, I think that's also perfectly reasonable (and would remove this question)

I'm leaving towards what you have now, at least in terms of functionality. Being able to do one-versus-rest type aggregations, perhaps over a neighborhood graph, could be cool, if it's supported i.e., vec of 1s and 0s (neighbors and not neighbors). Not sure if people will really do this though, but just a thought.

what do you think of the third option from my initial comment in this thread?

Oh missed this

don’t stringify at all and use AdRefs as df.columns?

Is this serializable? What if someone tries to write their results?

Hmm, the problem is that stringifying it introduces one more format that isn’t really supported. We have

1. the actual AdRef object
2. the simple string version (supported via A.resolve, e.g. "obs.some_col". Can only represent a subset of possible refs)
3. the repr (not supported as conversion format anywhere)

We could fix that by doing something like

• here, if passing a collection of AdRefs, serialize to repr always (instead of the conditional .idx we have now)
• in A.resolve, also support the repr format (if input.startswith('A.'): …)

then people can just use A.resolve to get back the input AdRefs from the obs column names.

Or we serialize/deserialize AdRefs in AnnData …

Does it make sense to check the names of the result i.e., result.obs_names?

Also the obs of result should contain a certain column derived from by, right?

yeah but: #4199 (comment)