Allow skipping loading all apigroups on startup

[lilatomic]

Which project are you requesting an enhancement for?

kr8s

What do you need?

As part of the startup of a kr8s Api, async_api_resources_uncached will make an API query for every apigroup in the cluster. If there are many clusters that are far away and have many api-groups this can take a few seconds. We have many small cli tools that leverage kr8s, so this loading time sometimes takes most of the execution time. It would be nice if we could bypass this.

A few thoughts on how we could proceed:

1. allow manually loading APIObjects
2. prime the cache with the results similar to _walk_subclasses(APIObject). This will provide a reasonable default
3. kubectl already caches the responses from these endpoints. We could load them directly

Having an incomplete cache should be fine, since we could fall through and do a full load. I think there is a risk in loading something that is possibly stale (such as the builtin APIObjects) since they could contain incorrect data. But we could mitigate that the way kubectl does.

I think we would need to expose the cache of async_api_resources. If you're OK modifying the vendored asyncache, that's a single-line change.

[jacobtomlinson]

Given the core goal of this library is to mimic kubectl behaviour I think persisting the cache to disk is the most aligned approach.

It would be interesting to see if we can reuse the kubectl cache, or whether we should have our own format.

[AlexGS74]

I would second this request. Kubectl is a generic tool, so it has to go broad and deep. When we use this library to create specialized tooling, that approach is slow and expensive oftern without giving any benefit. Me currently, I have to use call_api and strongly typed wrappers around it to bypass the discovery. it still does that version checking to establish the connection. Which I would also prefer to not have as unnecessary in my case. There is a difference between having opinions and preferences and not allowing extensions or deviations from those. This is a library in the end, not a tool so it should be natural for it to be mendable to fit into specific requirements or specific application cases.