PEP 688, take 2: Making the buffer protocol accessible in Python

I just posted a revised version of PEP 688, which now proposes a Python-level API for the buffer protocol, adding the new __buffer__ and __release_buffer__ dunders. This allows for buffer protocol support in the type system, and also allows Python classes to implement the buffer protocol.

The previous discussion was here: PEP 688: Making the buffer protocol accessible in Python

Possible topics of discussion:

  • The PEP still proposes removing the special case where “bytes” is supposed to also mean “bytearray” and “memoryview” in annotations. This proposal got some pushback in the previous thread.
  • Does the interaction between the C buffer API and the proposed Python API make sense?
  • A few opportunities for bikeshedding: where should the Buffer ABC live, what should the new special methods be called, should we add any other related useful tools to the standard library?

As this feature is a feature of the language, as well as of the C-API, is it possible to specify the Python behaviour without reference to the implementation language, then the change to the C-API and slots separately? I’m not arguing against discussing the reference implementation as evidence of feasibility.

My interest in this is not primarily typing (the category the PEP has).

The example is not enough for my understanding of the possibilities in general because it is a special case. Taking the memoryview in the constructor shape-locks the bytearray for the lifetime of the object by upping the export count, but suppose we wanted not to?

Let MyBuffer admit an append operation delegated to the bytearray, presumably only possible when the MyBuffer is not exporting. The __buffer__ method would have to return a fresh memoryview, and as long as any had not been returned to __release_buffer__, calling append would fail, right?

MyBuffer.__release_buffer__ would have to call memoryview.release. I think I understand this is not the same memoryview handed out, although it wraps the same Py_buffer.

It seems like, in this story, the MyBuffer is not conscious of its export count, since the bytearray takes care of it invisibly. It could track them if it wanted to by pairing off the calls, but not in a boolean as now.

Would this make a good mutable example?

Thanks for the feedback! I put up a PR at PEP 688: Enhancements from discussion by JelleZijlstra · Pull Request #2832 · python/peps · GitHub that

  • extends the specification to focus more on the behavior, not on the C slots
  • expands the MyBuffer example with an extend method similar to what you outline

Let me know if this addresses your concerns.

Yes, I think that’s better for starting with the Python behaviour.

It is challenging to follow what exactly is happening with wrapping in a memoryview. However, these specifications always need careful reading, and one’s interpretation to be checked with the reference implementation.

What I expect to see in a hypothetical reference slot_bf_buffer, for example, is code that begins like vectorcall_method(&_Py_ID(__buffer__), ...) and discards the memoryview wrapper it gets back. And in slot_bf_buffer_release I expect code that synthesises a new memoryview wrapper around the passed Py_buffer.

This doesn’t seem quite what is implied by:

When this method is invoked through the buffer API (for example, through memoryview.release), the passed memoryview is the same object as was returned by __buffer__.

The signatures of slot_bf_buffer and slot_bf_buffer_release are:

typedef int (*getbufferproc)(PyObject *, Py_buffer *, int);
typedef void (*releasebufferproc)(PyObject *, Py_buffer *);

so there is no opportunity to guarantee as you have in MyBuffer.__realease_buffer__ that:

    def __release_buffer__(self, view: memoryview) -> None:
        assert self.view is view  # guaranteed to be true

If I use the Python API, calling __buffer__and __release_buffer__ myself, then I have to pass exactly the memoryview I received, since I have no other way to be sure I an returning the same Py_buffer. But when memoryview.release does this to a buffer object defined in Python, it will call the slot containing slot_bf_buffer_release and the buffer object in Python will receive the new ephemeral memoryview.

Sorry for bad news, but:

if a buffer is writable, it must define the bf_releasebuffer slot, so that the buffer can be released when a consumer is done writing to it.

That’s not actually true. The bf_releasebuffer slot is needed when there’s extra cleanup to be done, other than decref(view->obj). That’s orthogonal to the data being writable (PyBUF_WRITABLE).
For example, NumPy arrays are mutable, but set bf_releasebuffer to NULL.
Types with mutable “shape”, like bytearray which has append, generally need to implement bf_releasebuffer so they can prevent resizing while a buffer is exported. But I don’t think “mutable shape” is a very useful property for this PEP, and anyway it’s not the only possible reason for implementing bf_releasebuffer.

I believe this came from reading the argument parsing docs? The wording there is unfortunate – by a stretch, one could argue it’s technically correct (if you read it as giving a specific local definition to the terms “mutable”/“read-only”), but it definitely suggests wrong stuff for uses outside arg parsing.

I guess tp_as_buffer could grow field(s) for “not supported” and/or “~guaranteed support” PyBUF* flags? At the moment it’s really only available at runtime by EAFP, as far as I know. Or perhaps it’s not useful for the C API, and could just be Python class attributes or some kind of typing-only hints.

Thanks for bringing this up! I got this idea from a post in the previous thread: PEP 688: Making the buffer protocol accessible in Python - #29 by storchaka. It took me a while to convince myself that it’s true, but maybe it’s not! The line of thinking that persuaded me was that anything that allows a writable buffer must keep track of whether someone is already writing to it in order to be thread-safe, and therefore it must have a bf_releasebuffer slot. Among the stdlib’s buffer classes, it also seems to be true that all the mutable ones have bf_releasebuffer slots. Numpy gets around this with manual refcounting: the buffer returned by bf_getbuffer holds a reference that the consumer must eventually DECREF, which signals to the numpy array that its consumer is gone.

If we can’t use bf_releasebuffer to signal mutability, I think we’ll have to go back to a single Buffer type, with no affordance for mutability in the type system. There just isn’t an elegant way to support it.

The current implementation does indeed provide this guarantee, and that’s why I am comfortable making this guarantee in the spec. The prototype is here: cpython/typeobject.c at de3a4bc518f7137a7a3c86cc7fc28b70fcb42013 · JelleZijlstra/cpython · GitHub

I implemented it this way because we need to ensure that when the consumer of the buffer created through __buffer__ releases it, we call __release_buffer__ on the same Python object. But if we just returned the buffer inside the memoryview we got from __buffer__, there’s no way to get back to the original object. So the implementation creates another wrapper that holds a reference to the memoryview and the Python object. When the buffer is released, we properly call __release_buffer__ while also cleaning up whatever is inside the memoryview.

One use-case where a bf_releasebuffer for a read-only buffer might come up is accessing buffers managed by some other system, for example when interfacing with a system that uses a moving GC.

Not the .obj of the Py_buffer being released? I know this can be NULL, in principle, but that’s when there isn’t meaningfully an underlying object. We are interested in the case of a C client, so it would call PyBuffer_Release(). That seems to manage ok given just a Py_buffer.

The guarantee seems to make this quite complicated, and I don’t see why it is needed.

Edit: Ok, I see the leap I’m making. The .buf of the buffer in the memoryview is the bytearray (in the example of use) not the MyBuffer instance itself. So you want a second layer to this where view->obj in PyBuffer_Release(Py_buffer *view) is the the MyBuffer instance.

Edit 2: Having got that clear, I think it is now difficult to see how an object would give you, through the __buffer__ method, a memoryview that truly represented a buffer view of itself, rather than only ever one bytes-like member of itself, of a built-in type that was already capable of of the buffer interface. Is it useful enough if it only does that?

I wonder if a good examples to illustrate the idea would be:

  1. objects that hold images in a compressed form, but will export their image as an array of decoded integer values.
  2. an audio equivalent of that.
  3. a sparse array that unpacks to an array whilst exported.

I looked at mmap as a possible example, and it expects to be released, whether read-only or not.


There’s no rule that objects must be thread-safe. Thread safety is a very nice property to have, sure, and I’d expect well-behaved objects to have it, but I’d also expect some objects that expose raw memory to eschew it in favor of performance.

Also, while "one writer or N readers” is a good way to ensure safety, it isn’t strictly necessary. (Think of a buffer of results, where each result can be set independently by a different thread. It might not be too performant, but it should work.)

Yes. That definitely contributes to the confusion.

All Python buffers hold that reference (view->obj). PyBuffer_Release DECREFs it. There’s nothing manual about it from NumPy’s point of view.

Going back to C argument parsing which I think is the source of this confusion:

  • PyArg_ParseTuple can take an object with bf_releasebuffer=NULL and return its buffer as const char*, because of PyArg_ParseTuple’s very special semantics: its inputs outlive its outputs, and so it doesn’t need an INCREF (i.e. the returned const char* has a hidden borrowed reference).
  • PyArg_ParseTuple can not take objects with bf_releasebuffer and convert them to const char*, because it has no way to call bf_releasebuffer when the buffer is no longer needed. To handle these types, the user must ask for Py_buffer (or a Python object).

So, simple C functions (ones that use PyArg_ParseTuple and deal with const char* internally) don’t allow types with bf_releasebuffer set, which means that they disallow common mutable types such as bytearray. It’s easy to read that as “mutable types have bf_releasebuffer set”, especially when there are no good counterexamples in the stdlib.

I decided to give up on distinguishing mutable vs. immutable buffers for now, especially because Introspection and "mutable XOR shared" semantics for PyBuffer may give us a more robust way to tell them apart in the future. This will make the PEP simpler but still useful for typing. A new version of the PEP just live at PEP 688 – Making the buffer protocol accessible in Python |

FYI, I filed gh-98712 to clarify the C arg parsing docs.

Unless some additional feedback comes up, I plan to submit the PEP to the SC in the next few weeks.

Over the last week I have been reviewing bytes type annotations in typeshed (Track review of `bytes` types · Issue #9006 · python/typeshed · GitHub); so far I have covered nearly all of the stdlib. This uncovered many more places where the standard library accepts bytes but not bytearray, which in my mind strengthens the case for dropping the implicit conversion between the two. I posted some tweaks to the PEP (PEP 688: Small tweaks by JelleZijlstra · Pull Request #2866 · python/peps · GitHub) to reflect that.


I haven’t reviewed __release_buffer__ yet. I’ll need to find a chunk of time to wrap my head around it again.
Currently it’s not clear to me when to call __release_buffer__ vs. when a wrapper or memoryview does it for you: e.g. the PEP says “It is also possible to call __release_buffer__ on a C class that implements bf_releasebuffer” – when should you do it, and what happens if you should do it but don’t (and vice versa)?

I took the reference implementation and added assert (self->ob_exports >= 0); to array_buffer_relbuf. That asserts a C-level invariant – it’s definitely not something I should be able to trigger through Python code. But I can:

>>> import array
>>> a = array.array('b', range(5))
>>> m = a.__buffer__(0)
>>> a.__release_buffer__(m)
>>> a.__release_buffer__(m)
python: ./Modules/arraymodule.c:2606: array_buffer_relbuf: Assertion `self->ob_exports >= 0' failed.

FWIW, when I get time I want to check if the following would be a better (simpler/safer) approach:

  • memoryview itself gets an extra field for “Python re-exporter”
  • A Python implementation of __buffer__ must return memoryview(..., exporter=self).
    • The __buffer__tp_getbuffer wrapper checks this
  • memorview calls the re-exporter’s __release_buffer__ on release() (unless released already).
  • The tp_releasebuffer__release_buffer__ wrapper only checks the argument and calls release() on it, thus only calling code that’s safe to call from Python.
  • You never need to call __release_buffer__ manually, but if you do, nothing bad happens (there’ll probably be recursion/reentrancy issues to solve in guaranteeing this)

You should call it when you want to release a memoryview you got from a call to __buffer__. It’s probably going to be very rarely necessary, because you can just do with obj.__buffer__(flags) and the memoryview’s __exit__ method will call the buffer for you. Similarly, if you don’t ever call it, the runtime will release the buffer for you when the memoryview you got from __buffer__ gets GCed. I will add some discussion of this to the PEP.

Good catch! I’m testing a change to the reference implementation now that basically calls memoryview.release from wrap_releasebuffer, so that memoryview takes care of checking that we’re not re-releasing an already released buffer.

This isn’t too different from how the reference implementation already works. However, instead of setting a flag on the memoryview returned from __buffer__, I wrap it in another memoryview that remembers its Python exporter by setting its obj to a wrapper type that references the Python exporter and the underlying memoryview that __buffer__ returned. This avoids changes to memoryview:

>>> class X:
...     def __buffer__(self, flags): return memoryview(b"x")
>>> mv = memoryview(

>>> mv = memoryview(X())
>>> mv.obj
<_buffer_wrapper object at 0x10120e670>
>>> import gc
>>> gc.get_referents(mv.obj)
[<memory at 0x10114cd50>, <__main__.X object at 0x1011ab990>]

How important is that, really?

It won’t:

>>> class Foo:
...     def __buffer__(self, flags):
...         print('__buffer__ called')
...         return memoryview(b"here's your data")
...     def __release_buffer__(self, view):
...         print('important releasing code here')
...         view.release()
>>> with Foo().__buffer__(0):
...     pass
__buffer__ called
>>> # release not called!

Hence the idea of requiring memoryview(..., exporter=self) in def __buffer__.

I’d be open to adding new Python APIs if that’s helpful, but every new Python API has some cost (documentation, bikeshedding over naming/semantics, having to reason about correctness if it’s used in unexpected ways). So I’d prefer to avoid introducing a new Python API if possible.

If you call __buffer__ directly, you’re on your own. I see this as similar to the context manager protocol: __enter__ and __exit__ should be called in pairs, but if you call __enter__ directly and don’t call __exit__, the runtime isn’t going to call it for you.

If you use with memoryview(Foo()): instead, Foo.__release_buffer__ will be called.

Such a requirement wouldn’t be enforceable in code that directly calls __buffer__ (at least, not without bigger changes to the language).