allennlp (0.7.2) depends on cffi (1.11.2), gevent (1.3.6) is incompatible wi th allennlp (0.7.2).

[phiresky]

this issue still happens (or happens again) with allennlp v0.7.2

[SolverProblemError]                                                  
Because gevent (1.3.6) depends on cffi (>=1.11.5)                                  
 and allennlp (0.7.2) depends on cffi (1.11.2), gevent (1.3.6) is incompatible wi  
th allennlp (0.7.2).                                                               
And because allennlp (0.7.2) depends on gevent (1.3.6), allennlp is forbidden.     
So, because no versions of allennlp match >0.7.2,<0.8.0                            
 and code depends on allennlp (^0.7.2), version solving failed.

Originally posted by @phiresky in #1824 (comment)

[schmmd]

@phiresky I'm unfamiliar with poetry. Can you explain to me how I can repeat the error you are seeing on my own laptop?

[phiresky]

Sorry, I don't have a computer right now but this should work: pip install poetry mkdir x && cd x poetry init -n poetry add allennlp (Might have to change the min py version in pyproject.toml to ^3.6.1)

…

On Wed, Dec 19, 2018, 20:43 Michael Schmitz ***@***.*** wrote: @phiresky <https://github.com/phiresky> I'm unfamiliar with poetry. Can you explain to me how I can repeat the error you are seeing on my own laptop? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2195 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACMnYbTz4r2VPT0JTkDfYpjO_1cTdYVbks5u6pbggaJpZM4ZWjAG> .

[sodre]

We experienced the same problem when building the AllenNLP package for conda-forge/staged-recipes#7324. Our solution to that problem was to relax the pinning on cffi to 1.11.*.

This brings me to a slightly larger question of whether the pinnings in AllenNLP are too strict. This issue came up during the peer-review process that every new conda-forge package goes through. Please see comments 1 and 2.

cc: @bgruening

[schmmd]

Closing, since we have a new issue with 0.8.0 (#771).

[sodre]

Sorry, but how is this issue related to #771? 771 is about codecov and this one is about package pinning.

[schmmd]

Sorry @sodre--I got confused when I followed the link to python-poetry/poetry#771. It seems like this is still a problem for AllenNLP 0.8.0 and we should relax our bounds on cffi. I'm confused however on how relaxing our bounds on cffi fixes the issue since @phiresky reports that gevent (1.3.6) requires cffi (>=1.11.5) and AllenNLP directly depends on gevent==1.3.6.

[sodre]

No problem.

By relaxing allennlp's requirement on cffi to 0.11.* instead of 0.11.2 while keeping gevent at 1.3.6 allows the solver to pick cffi 0.11.5 to solve the package dependency graph.

[schmmd]

I'm confused because in the requirements.txt and setup.py we pin cffi to 0.11.5. Your last comment and the title of this issue claims AllenNLP pins cffi to 0.11.2.

Clearly, I'm not expert on Python dependencies.

@sodre do you suggest a change such as https://github.com/allenai/allennlp/pull/2271/files?

[sodre]

@schmmd, The original issue was for AllenNLP 0.7.2, which had it pinned at 0.11.2.

allennlp/setup.py

Line 112 in bf04f1d

'cffi==1.11.2',

As far as a suggestion, here is what gevent uses.

# The exception is on Windows, where we want the libuv backend we distribute
# to be the default, and that requires cffi; but don't try to install it
# on PyPy or it messes up the build
cffi_requires = [
    "cffi >= 1.11.5 ; sys_platform == 'win32' and platform_python_implementation == 'CPython'",
]

I have not tried to do a build against 0.8.0 yet in conda forge. So I am not sure if the error would still happen with your latest setup.py, even without #2271. My gut tells me it should work just fine.

[schmmd]

Since we're not even sure this is a problem with the latest release (0.8.0) I'm going to close this issue in favor of the one I just opened (#2280) about creating a continuous integration check for dependency conflicts. I'd love suggestions about what tool we could use to do such a check.