Consider allowing a `CollectionNotStr` rather than `SequenceNotStr` for column spec.

[randolf-scholz]

Afaik, pandas to this day lacks the ability to directly construct a DataFrame with a specific schema. (pandas-dev/pandas#4464)

The most direct way I know of to get a DataFrame of a desired schema, given we already have a raw data object, is:

import pandas as pd
data = [(1, 1.2, "a"), (2, 2.3, "b")]
schema = {"id": int, "value": float, "name": str}
df = pd.DataFrame(data, columns=schema).astype(schema)

This works at runtime but flags with type-checking, because schema is not SequenceNotStr, requiring us to use a runtime-redundant list(schema) instead.

Proposal: introduce a CollectionNotStr, which would match dicts like schema and use that instead of SequnceNotStr where appropriate.

from typing import Protocol, Iterator


class CollectionNotStr[T](Protocol):
    def __iter__(self) -> Iterator[T]: ...
    def __len__(self) -> int: ...
    def __contains__(self, value: object, /) -> bool: ...
    
x: CollectionNotStr = {"id": int, "value": float}  # OK
y: CollectionNotStr = "sas"  # ERROR

This affects most methods that accept a columns=... parameter.

[Dr-Irv]

This is a documentation issue within pandas that has to be dealt with first, so I suggest you create an issue there to generate a discussion about whether we want to support a dict as the columns (or index) parameter to DataFrame, and even to Index. It's either a doc change OR code should be added to reject a dict as the argument.

For DataFrame, the index or columns arguments say "Index or array-like". We don't consider a dict to be either. Similarly, for pandas.Index, the docs say that data should be "array-like (1-dimensional)". Yet pd.Index({"a":1, "b":2}) works, which is why it works when using a dict for either the index or columns arguments in DataFrame.

For pandas-stubs, we follow the docs to determine what is allowed.

[Dr-Irv]

Discussed in pandas dev meeting 8/12/26. Created issue to update the docs:
pandas-dev/pandas#66742

Can update the stubs to allow Mapping[HashableT, Any] to accommodate changes there.