get_all_specs not compatible with subclassing

[minrk]

nb_conda_kernels subclasses KernelSpecManager to find conda kernels. It does so by overriding find_kernel_specs to return its name list and resource directories, and get_kernel_spec to return the specs themselves.

We've added .get_all_specs as a simpler API to get all of the specs without separately finding and loading them. But we've done this in such a way that .get_all_specs calls .find_kernel_specs, a public API, and _get_kernel_spec_by_name, a private method. This means that implementations that override .find_kernel_specs in such a way that _get_kernel_spec_by_name won't work, then .get_all_specs is broken for any such subclass (this includes nb_conda_kernels 2.1).

Now, one option is for custom kernelspec managers to implement all three methods - find_kernel_specs, get_all_specs, and get_kernel_spec, but this is a backward-incompatible change, and I think we should consider it a bug.

I think we should ensure that .get_all_specs works for KSM subclasses that have only implemented the public methods .find_kernel_specs and .get_kernel_spec to fix this backward-incompatibility, and push that out in a bugfix release.

cc @dartdog who reported the issue

[michaelaye]

I'm puzzled why things break with conda 4.4.x while they still work with 4.3.x? Where's the conda relationship here?

[minrk]

@michaelaye shouldn't be a sensitivity to conda version. The relevant version change is notebook 5.3 which started using the get_all_specs method, revealing this bug.

[escorciav]

Is this still a problem with jupyter-notebook==5.4.1?
I am using nb_conda_kernels==2.1.0 from conda-forge but I can't find another environment.
I double check that ipykernel is in the other environment.

[takluyver]

@escorciav you need to make sure you have jupyter_client version 5.2.2 or above.

[escorciav]

thanks for the heads-up @takluyver
My conda environment reports jupyter_client 5.2.3 py36_0. Any other clue?
I'm using conda 4.5.1.

[takluyver]

If you have multiple envs, check which one the notebook server is running from, and the version of jupyter_client in there. If that's the newest version, then it's not this bug - try asking nb_conda_kernels about it.

[escorciav]

Thanks. Everything seems fine in terms of package version. Although, the default kernel is called python3 rather than the name of the environment.

• Are you using nb_conda_kernels?
  If yes, could you tell me the version of jupyter_client, jupyter_lab, nb_conda_kernels and conda that u have?

[takluyver]

I'm not using nb_conda_kernels. I confess I don't actually use environments very much - I tend to install everything with --user.