Repository navigation
Support for PEP 567 #420
Description
PEP 567 got accepted on the 22nd, adding context-local variables to Python 3.7, and it would be super useful to have implemented (especially for something I'm working on).
However, 3.5/3.6 compat is an issue, and if it wasn't for that I would've PR'd in already (since the change is probably super trivial, msg = task.coro.send(next_send) -> context = contextvars.copy_context(); msg = context.run(task.coro.send, next_send) or similar.), and I don't know how that would be solved.
Activity
njsmith commentedon Jan 24, 2018
Yeah,
ContextVaris super cool!What I'm thinking we should do is deprecate
trio.TaskLocal, and instead have atrio.ContextVarclass. On 3.7 this will just befrom contextvars import ContextVar, and on earlier versions it'll be our own class that's designed to have the same semantics, but implemented using the same techniques that we currently use forTaskLocal.Unfortunately, we don't currently have any way to test on a recent 3.7 snapshot (travis-ci/travis-ci#9119), so probably that will need to be resolved first. In particular, we'll need to write tests for our workalike implementation, and I want to run those tests against the real thing too to make sure we don't accidentally mess up the semantics!
(BTW, just in case you haven't seen
TaskLocal, you might want to check it out, since it does work now.)njsmith commentedon Jan 24, 2018
Oh, and we'll want to be calling
copy_contextfrom_spawn_impland attaching the context to the task object, and then doingtask.context.run(task.coro.send, next_send).njsmith commentedon Jan 24, 2018
It looks like Travis's
3.7-devandnightlyPythons are now updated to have PEP 567 support, so that's no longer a blocker.It did occur to me that for supporting older Pythons, then instead of having
trio.ContextVar, another option would be to have a backport library (pip install pep567orpip install contextvarsor whatever), and then use it in Trio. In some sense this would be equivalent: there'd be a library on PyPI that contains code to emulate thecontextvarsAPI. But by splitting that off into a separate library from trio, it would potentially be usable by other projects too.Personally I'd be happy with a version of this library that was in pure Python and didn't bother monkeypatching asyncio, but I think @1st1 has been thinking about a more ambitious backport project, so CC'ing him to get his thoughts.
m-u-xyz commentedon Feb 15, 2018
pypi doesn't have
pep567, andcontextvarsis incomatible with PEP 567.So the first step would be to implement the "real"
contextvarsin pure Python.We don't need to monkeypatch asyncio, since trio-asyncio will supersede the affected functions anyway.