WIP: raise an error when using a lambda and default executor

[basnijholt]

Closes#206.

It is not obvious that this blocks the kernel, but it does hang forever and does not error.

importadaptiveadaptive.notebook_extension()
f=lambdaxy: xy[0]**2+xy[1]**2learner=adaptive.Learner2D(f, bounds=[(-1, 1), (-1, 1)])
runner=adaptive.Runner(learner, goal=lambdal: l.npoints>=1000)

This PR makes sure an error is raised, however, might this be caused by an underlying issue?

[akhmerov]

Here's my thinking:

• Does not work with lambda functions #206 is likely not a pressing issue
• There are other other common cases where pickling fails (e.g. closures). Those aren't going to be caught by a type check.

Therefore I propose to not apply a patch like this, but rather wait until the issue becomes more important, or someone (maybe one of us) is motivated to have a proper fix.

[basnijholt]

To be fair, two persons now reported this issue separately.

See #211 for the other one.

[akhmerov]

Perhaps a better implementation would be to try to pickle the function on our own and see if that raises?

[basnijholt]

I have added that check.

I still also explicitly check whether it is a lambda because that is the most common case I think. Instead, we could probably also just add that in the error message.

[jbweston]

Just that minor clarification in the error message to make it very clear what users need to do to make it go away.

Other than that LGTM

[jbweston]

this check doesn't actually check if the function is a lambda. Check it out:

import types
def f():
pass
print(isinstance(f, types.LambdaType))

prints True

[jbweston]

This is why the tests are failing AFAICT

[basnijholt]

I've made the check a bit simpler.

[basnijholt]

I actually ran into this issue myself today!

I also simplified the test a bit. @jbweston or @akhmerov should I rephrase the error message or merge as is?

[jbweston]

The tests still fail and the failure is relevant to this issue.

AsyncRunner does not actually use the executor if the function passed in is an async def function. In this case the (async def) function is not pickled at all and we don't need to worry about pickling.

The check should be changed to

if executor is None and not inspect.iscoroutinefunction(learner.function):

[basnijholt]

Kim just ran into this problem too.

[jbweston]

If we fix the test failures (i.e. fix the code) we can merge!

[jbweston]

LGTM