Provide a way to drop resources in Python Land

[erikrose]

In some instances, it is necessary to promptly drop resources: for example, to grab a second request for Fastly's session-reuse feature. Currently, we use an ugly hack:

# Here we abuse an arbitrary request-consuming function to trigger
# the drop. Glue code interposed by wasmtime's linker ensures that
# drop happens, but send() otherwise fails before doing anything.
send(request, body, "no such backend")

Without this, a trap occurs at await_request(), IIRC. (If I don't recall correctly, it's at next_request().)

I see 2 ways out of this:

1. Convince the Python runtime to drop things that Python itself GCs (either because refcount drops to 0 or because the cyclic GC ditches them). This would enable us to implement explicit drops in our SDK by simply saying del some_resource behind nice abstractions. Customers could also call del if they wanted to program at a lower level. Most of the time, I imagine we could do nothing and let scope exits take care of dropping things.
2. If that proves difficult for some reason, we could generate a drop_foo() (likely in componentize-py) for each resource type foo and call those instead of del.

Open question:

• How are drops currently effected? Are they at all?

[erikrose]

2 thoughts:

1. del may not cut it: it certainly doesn't guarantee the refcount goes to 0, and we're talking about crashes here if it doesn't. A drop_request(the_request) could be easily called from a close() (or whatever) disposal method, keeping everything reliable no matter what tangled reference graph you weave.
2. See how JS does it, since its situation is very similar to Python's (minus refcounting).

[erikrose]

Our JS SDK does explicitly and deterministically call whatever_type_drop_own(), so it seems to agree with number 1 above.

[erikrose]

There is a suspicious "Release this resource" docstring applied to generated __exit__() methods on resource types. In our current tree, this just spits out raise NotImplementedError. However, I see that's under an if stub_runtime_calls branch. The alternate branch actually appears to dispose of things. Maybe we just need to turn stub_runtime_calls off and start using resources as context managers.

[erikrose]

Joel says drop is supposed to work, either by contextmgr's __exit__ on the value or when its refcount goes to 0.