You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In #133 there was a request to be able to provide a custom _HookCaller (in this case in order to be able to warn on deprecated hook calls - since addressed via #138).
This feature may be quite useful in the future not only in its ability to allow users to customize the pluggy runtime but also if we want to eventually support async / coroutine based plugins (this can be done loop agnostic these days with something like multio and/or trio-asyncio).
I'd like to get feedback on this.
To start it would mean that we would make a simple change where PluginManager can take a hook_caller_type as input during instantiation.
How about multiple hook callers? At work we might need to use a pluggy hook call system identical to pluggy, but also capable of loading hook implementations from shared c++ libraries.
It's already possible to define a synchronous hook that internally does trio.run(something_async) or asyncio.run(something_async). To get extra value from async hooks, I think you'd need to somehow "expose" the async-ness to the caller, so that the caller can be async, and do things like:
so that multiple hooks are running under the same event loop.
In principle I'm guessing this wouldn't be too complicated. But there's not much point unless one of pluggy's users (like pytest) is interested in using it in this mode.
@njsmith I've toyed with this idea and what I ended coming up with is that hookimpls essentially become coroutine factories and the caller awaits their results, that way you can call from it inside another coroutine without pissing off the event loop (not sure how trio or curio handle the equivalent of loop.run_until_complete from inside a running coroutine, but I imagine they don't take kindly to it).
I never expanded it beyond toy examples so I'm not sure what shortcomings I missed beyond the caller needs to reimplement some of pluggy's filtering logic.
Hi, I maintain a feed reader library that supports plugins, and even if using pluggy would require some work-arounds for calling, I think the collection / discovery functionality is still more than worth it.
I want to describe the use cases I have in this comment, since it may help inform the design for custom hook callers.
In theory I think there are work-arounds for all cases, and with a bit of effort on the caller's part, without affecting plugin writers.
for (2), wrap HookImpl.function to append exceptions to a list
for (3), pass a reference to the request, and instead of returning a value, update the same reference; again, this can be done by a wrapper (decorator kind, not pluggy kind) over the original hook function (but accepting fewer arguments may need to be re-implemented in the wrapper)
(4) is a more complicated version of (3); presumably the "set reference" logic can also decide whether to re-run the action or not
AFAICT, supporting hook wrapper versions of the above may prove difficult, both in pluggy and in the work-around (lots of semantics to be decided).
Activity
RonnyPfannschmidt commentedon May 22, 2018
i propose deferring this until we can experiment against trio and asyncio
based on initial example i'd absolutely avoid multio
RonnyPfannschmidt commentedon May 22, 2018
oh - we might want to investigate removing multi-call by moving it into a hook-caller subclass/mixin
nicoddemus commentedon Jun 5, 2018
How about multiple hook callers? At work we might need to use a pluggy hook call system identical to
pluggy, but also capable of loading hook implementations from shared c++ libraries.cc @williamjamir
goodboy commentedon Jun 5, 2018
Love both of those ideas!
RonnyPfannschmidt commentedon Jun 5, 2018
@nicoddemus sounds like overkill a hook parser or a wrapper sounds more viable there
njsmith commentedon Jun 13, 2018
It's already possible to define a synchronous hook that internally does
trio.run(something_async)orasyncio.run(something_async). To get extra value from async hooks, I think you'd need to somehow "expose" the async-ness to the caller, so that the caller can be async, and do things like:so that multiple hooks are running under the same event loop.
In principle I'm guessing this wouldn't be too complicated. But there's not much point unless one of pluggy's users (like pytest) is interested in using it in this mode.
justanr commentedon Jun 13, 2018
@njsmith I've toyed with this idea and what I ended coming up with is that hookimpls essentially become coroutine factories and the caller awaits their results, that way you can call from it inside another coroutine without pissing off the event loop (not sure how trio or curio handle the equivalent of
loop.run_until_completefrom inside a running coroutine, but I imagine they don't take kindly to it).I never expanded it beyond toy examples so I'm not sure what shortcomings I missed beyond the caller needs to reimplement some of pluggy's filtering logic.
RonnyPfannschmidt commentedon Aug 16, 2021
i had a sync with @simonw on the weekend and i will try to incooperate the details into pluggy as i sprint on speedups this week
lemon24 commentedon Oct 18, 2024
Hi, I maintain a feed reader library that supports plugins, and even if using pluggy would require some work-arounds for calling, I think the collection / discovery functionality is still more than worth it.
I want to describe the use cases I have in this comment, since it may help inform the design for custom hook callers.
before feed update hook – first exception bubbles up; already supported
after feed update hook – collect exceptions and raise as a group; Feature Suggestion: Collect multiple exceptions from hook implementations #419
request hook – hook return value can replace one of the arguments
response hook – maybe repeat action with hook return value
In theory I think there are work-arounds for all cases, and with a bit of effort on the caller's part, without affecting plugin writers.
AFAICT, supporting hook wrapper versions of the above may prove difficult, both in pluggy and in the work-around (lots of semantics to be decided).