Enable unlisted_choice in LEAP hub
#3047
Conversation
|
Merging this PR will trigger the following deployment actions. Support and Staging deployments
Production deployments
|
yuvipanda
force-pushed
the
unlisted-choice-take-2
branch
from
3d256f9 to
8d0d3d1
Compare
|
This unfortunately still has the problem previously reported, where once you use an image as the other choice, you can't use any other image :( This shouldn't be deployed to the LEAP prod hub until we figure that out. |
I reproduce this once before in staging.leap.2i2c.cloud earlier today, but I've not been able to reproduce it later or with a local k3s cluster with jupyterhub 4.0.2 + kubespawner's main branch. |
|
I've just pushed a commit that I think fixes it. I've manually deployed it to staging https://staging.leap.2i2c.cloud and seems to be working. It looks like it was the same bug, with the same fix as the one in jupyterhub/kubespawner#766. But because we were using a callable |
|
Yesss, I can confirm that fixes it @GeorgianaElena!!!! \o/ |
|
🎉🎉🎉🎉 Monitor the deployment of the hubs here 👉 https://github.com/2i2c-org/infrastructure/actions/runs/6013781012 |
|
Nice work @GeorgianaElena!!! I think there is another fix to make in kubespawner also. I think this logic shouldn't just take the value returned by profile_list, but should make a copy of the value returned by a profile_list callable. if callable(self.profile_list):
profile_list = await maybe_future(self.profile_list(self))
else:
# Use a copy of the self.profile_list dict,
# otherwise we might unintentionally modify it
profile_list = copy.deepcopy(self.profile_list)
profile_list = self._populate_profile_list_defaults(profile_list)This upstream fix isn't a problem if the callable profile_list returns a copy at all times as done in the leap hub - but we don't return a copy at all times, for example when we return this: |
|
I'll submit a kubespawner PR to make a copy of the returned value of a profile_list callable, and then we should be fine for all hubs - but we will need the fix in b62a46a as well I think. |
|
@consideRatio I just looked into that, and I don't think that works - the problem is in , where we are taking the original value ofprofile_list, and modifying that is the problem. In the kubespawner code, the value returned by the callable isn't re-assigned to self.profile_list anywhere I can see, so is probably fine.
I think we should provide guidance that says "if you are using profile_list as a callable, make sure to make a copy!". But curious to see the outcome of your investigation too. Perhaps I'm wrong. |
I agree that the problem we have for the leap hub is the specific one patched by @GeorgianaElena, and would be a problem even if we patched what I consider patching in KubeSpawner. But, I think there is another bug, and I think we are exposed to it because we break the advice of "if you are using profile_list as a callable, make sure to make a copy!".
|
|
Hmm, I don't fully understand so am curious to see your PR @consideRatio. |
Ref #2146
Just brings in @GeorgianaElena's work in jupyterhub/kubespawner#769