collections.abc.ByteString is not equivalent to typing.ByteString

[hauntsaninja]

There are two related, but different strands of conversation here:

• collections.abc.ByteString is useless and confusing: Deprecate and schedule removal of collections.abc.ByteString and typing.ByteString #91896
• That bytes can be used as a shorthand for typing.ByteString and includes memoryview and bytearray https://peps.python.org/pep-0688/

However, there is an additional issue! collections.abc.ByteString is not an accurate replacement for typing.ByteString! collections.abc.ByteString only registers bytes and bytearray, whereas typing.ByteString is documented as also representing a memoryview. This is an issue regardless of the above two related conversations.

For this issue we could:

• Move the documentation of typing.ByteString out from under the "Corresponding to collections in collections.abc" section to the "Other concrete types" section
• Remove all mention of collections.abc.ByteString from typing.rst
• Keep the mention of typing.ByteString as being deprecated, but change the reason. ByteString is not a generic type and collections.abc.ByteString is not a semantic replacement for it as above, so the current reason is wrong on two counts
• Define typing.ByteString in typeshed as a simple Union, and not the Sequence[int] thing it is right now.
• Define typing.ByteString as a Union in typing.py instead of the generic alias it is now(?)

Here's how this relates to the other two strands of conversation:

• We should also remove the note about the bytes shorthand from the docs
• We should deprecate both typing.ByteString and collections.abc.ByteString. I'm fine with going slow on the removal of typing.ByteString, since once it's a union in typeshed it's not causing much harm. But ideally we can remove collections.abc.ByteString in 3.14, since its isinstance behaviour is not what you want.

cc @JelleZijlstra

[JelleZijlstra]

Thanks for the research here! Before we make changes in CPython it might make sense to wait for PEP 688 to be accepted (hopefully soon?), since that will affect what we should point users to.

I'm not sure I see the issue, though. At runtime typing.ByteString really is an alias for collections.abc.ByteString: ByteString = _alias(collections.abc.ByteString, 0) # Not generic.

And memoryview is not a subclass of either type at runtime:

In [51]: issubclass(memoryview, collections.abc.ByteString)
Out[51]: False

In [52]: issubclass(memoryview, typing.ByteString)
Out[52]: False

So while I am in favor of deprecating both ByteString objects, I don't see why we need to decouple them.

[hauntsaninja]

... I don't see why we need to decouple them.

I was hoping we could get away with treating typing.ByteString as bytes | bytearray | memoryview. This seemed like too breaking of a change for a collections.abc.ByteString, since the whole point of ABCs is to make your own subtypes of them and there are indications that collections.abc.ByteString was genuinely meant to be like bytes | bytearray.

I realise I was also kind of hoping we could just... pretend that collections.abc.ByteString doesn't exist, since I'd like to avoid users thinking there's an interface there.

Anyway, the claim is that typing.ByteString has clearly defined, reasonably well documented semantics here:

This type represents the types bytes, bytearray, and memoryview of byte sequences.

And while PEP 688 provides a path for something with better semantics (since users of typing.ByteString would probably want all buffers), maybe the existing semantics are still useful, and we can "fix" this outright bug in the implementation of typing.ByteString.

---

So possible "fixes" are:

1. Make typing.ByteString equivalent to bytes | bytearray | memoryview. As I type this, I'm realising this has the downside it would break any user who was subclassing typing.ByteString

(python/typeshed#9783 doesn't surface issues and I couldn't find any typing.ByteString subclasses on grep.app, and I found a collections.abc.ByteString one)

2. Register memoryview with collections.abc.ByteString. While desirable for use via typing.ByteString, there are indications that collections.abc.ByteString was genuinely meant to be just bytes | bytearray and e.g. have startswith and things like that. In which case this could break user expectations. E.g., the docstring says:

This unifies bytes and bytearray. XXX Should add all their methods.

edit: it is also tested that memoryview is not a collections.abc.ByteString

3. Don't "fix", and just aggressively pursue the deprecations.

[hauntsaninja]

I guess 2 might be more viable than I initially gave it credit for. This table is pretty clear that the interface for collections.abc.ByteString is __getitem__ + __len__. More authoritative than the "XXX" docstring, at least :-)

We can still go with option 3 and just deprecate instead of touching these, if we feel it's too messy.

[JelleZijlstra]

So your claim is essentially that there is a discrepancy between the implementation and the documentation for typing.ByteString, and we should fix that discrepancy by adjusting the implementation. I'm not sure why you don't want to change the documentation instead, though. memoryview is not a ByteString at runtime or according to typeshed, so it seems less disruptive to adjust the documentation.

My preference is to deprecate both variants of ByteString. As I see it, there are three possible use cases for ByteString, and it doesn't work well for any of them:

• A shorthand for "any buffer type". ByteString doesn't actually mean this; PEP 688 will provide a solid solution for this use case.
• A shorthand for "bytes | bytearray", or possibly "bytes | bytearray | memoryview" depending on which part of the docs we believe. Users who want this should just write "bytes | bytearray": it's barely any longer and does not suffer from the ambiguity of ByteString.
• An ABC representing the common interface of bytes and bytearray, as suggested by the "XXX" comment in the docstring. But since collections.abc.ByteString doesn't define any of this interface, it's not actually useful for this purpose. And as @rhettinger explained when I tried to add the shared bytes/bytearray methods, backward compatibility means we can essentially never add methods to an ABC.

ByteString is purely a source of confusion and I don't think it is useful. Let's deprecate it and remove it in 3.14.

[hauntsaninja]

Okay, great. #102096 adds deprecation warnings to collections.abc.ByteString

[srittau]

+1 on the deprecation and eventual removal. Since the situation is so muddled, and ByteString seems mostly unused, it's better to start from the beginning.