Closed astrojuanlu closed 1 year ago
cc @kirbyherm, let's see how this turns out.
Indeed, master
is currently broken.
@astrojuanlu @kirbyherm I think this is an issue triggered by an update in dill
. I'll investigate over the weekend.
Indeed, looks like dill<0.3.6
fixed the problems on 3.8 at least.
@astrojuanlu Thanks for checking this out. Not sure what the best course of action is.
IIRC dill
support was originally introduced after a user reported that cloudpickle
(the default serialisation backend in pygmo) was failing on some types. Thus I added the option to switch to dill
as an alternative. I don't think the selectable serialisation backend feature is being employed by many users, so my first instinct would be to just disable the dill
testing in the test suite.
The alternative is to chase this with upstream...
@astrojuanlu @bluescarni
I made a brief attempt to trace the errors back through dill
and quickly got lost. In #115 I employ @astrojuanlu 's solution of just reverting to dill=0.3.5.1
in all the ci/circleci testing environments. All the tests pass there.
Perhaps specifying the prerequisite for dill<0.3.6
might be a simple enough solution for now?
The proper solution would indeed be trying to report this upstream, but looks like a massive rabbit hole I don't have time to get in. Temporarily capping dill
at <0.3.6
for the time being so at least the CI can pass can be a good short-term solution.
As mentioned in #115, let's just pin dill for the time being. I'll be closing this one.
See if gh-112 and gh-113 CI failures are reproducible in current
master
.