Performance issue in Async Middleware handling

[Mykhailo Havelia]

Follow-up to #37177, for the part that was deferred there.

#37177 fixed per-hook context switching by setting async_capable = False on Django's built-in middleware. The cause is still in MiddlewareMixin: it declares async_capable = True by default, so any subclass with sync hooks goes through __acall__(), which wraps each hook in its own sync_to_async().

This defeats load_middleware(), which is designed to adapt once per contiguous block of sync middleware.

So every user and third-party middleware written the documented way (subclass MiddlewareMixin, implement sync process_request / process_response) still has the problem #37177 measured. A default project that adds one or two such middleware between the built-in ones also splits the sync block, adding more switches around them.

A side effect: these switches are invisible. __acall__() calls sync_to_async() itself, so adapt_method_mode() never logs "Asynchronous handler adapted for middleware ...", which is the tool the docs point users to for finding this.

Proposal

https://code.djangoproject.com/ticket/37177#comment:7

Deprecate the default MiddlewareMixin.__acall__() implementation: raise a deprecation warning when a subclass ends up using it with sync hooks (i.e. sets async_capable = True explicitly but doesn't override __acall__()). After the deprecation period, remove it and make async_capable = False the default.