On 01/11/2018 02:52, David Allemang wrote:

I do not think there is currently a good way for Context Managers to support suspended execution, as in await or yield. Both of these instructions cause the interpreter to leave the with block, yet no indication of this (temporary) exit or subsequent re-entrance is given to the context manager. If the intent of a Context Manager is to say "no matter how this block is entered or exited, the context will be correctly maintained", then this needs to be possible.

I think you're going to have to justify this a bit more. From my point of view, yielding does not leave the with block in any meaningful sense. Indeed I'd be quite hacked off with a file context manager that was so inefficient as to close the file on yielding a line, only to have to re-open and seek when it got control back. -- Rhodri James *-* Kynesim Ltd

Check out the decimal example here: https://www.python.org/dev/peps/pep-0568/ (PEP 568 is deferred, but PEP 567 is implemented in Python 3.7). Those Contexts aren't context managers, but still there's some thought put into swapping contexts out at the boundaries of generators. On Wed, Oct 31, 2018 at 7:54 PM David Allemang wrote:

I do not think there is currently a good way for Context Managers to support suspended execution, as in await or yield. Both of these instructions cause the interpreter to leave the with block, yet no indication of this (temporary) exit or subsequent re-entrance is given to the context manager. If the intent of a Context Manager is to say "no matter how this block is entered or exited, the context will be correctly maintained", then this needs to be possible.

I would propose magic methods __suspend__ and __resume__ as companions to the existing __enter__ and __exit__ methods (and their async variants). __suspend__, if present, would be called upon suspending execution on an await or yield statement, and __resume__, if present, would be called when execution is resumed. If __suspend__ or __resume__ are not present then nothing should be done, so that the behavior of existing context managers is preserved.

Here is an example demonstrating the issue with await: https://gist.github.com/allemangD/bba8dc2d059310623f752ebf65bb6cdc and one with yield: https://gist.github.com/allemangD/f2534f16d3a0c642c2cdc02c544e854f

The context manager used is clearly not thread-safe, and I'm not actually sure how to approach a thread-safe implementation with the proposed __suspend__ and __resume__ - but I don't believe that introducing these new methods would create any issues that aren't already present with __enter__ and __exit__.

It's worth noting that the context manager used in those examples is, essentially, identical contextlib's redirect_stdout and decimal's localcontext managers. Any context manager such as these which modify global state or the behavior of global functions would benefit from this. It may also make sense to, for example, have the __suspend__ method on file objects flush buffers without closing the file, similar to their current __exit__ behavior, but I'm unsure what impact this would have on performance.

It is important, though, that yield and await not use __enter__ or __exit__, as not all context-managers are reusable. I'm unsure what the best term would be to describe this type of context, as the documentation for contextlib already gives a different definition for "reentrant" - I would then call them "suspendable" contexts. It would make sense to have an @suspendable decorator, probably in contextlib, to indicate that a context manager can use __enter__ and __exit__ methods rather than __suspend__ and __resume__. All it would need to do is define __suspend__ to call __enter__() and __resume__ to call __exit__(None, None, None).

It is also important, since __suspend__ and __resume__ would be called after a context is entered but before it is exited, that __suspend__ not accept any parameters and that __resume__ not use its return value. __suspend__ could not be triggered by an exception, only by a yield or await, and __resume__ could not have its return value named with as.

Thanks,

David _______________________________________________ Python-ideas mailing list Python-ideas@python.org https://mail.python.org/mailman/listinfo/python-ideas Code of Conduct: http://python.org/psf/codeofconduct/

-- --Guido van Rossum (python.org/~guido)