Upgrade to the latest OpenAPI generator

[tomplus]

I've tried to use the latest version of OpenAPI generator to generate the library. That version uses Pydantic2 to generate models. Unfortunately kubernetes library contains ~600 models/api and loading time (at startup) is huge and not acceptable, especially for scripts, long running applications import modules once so it's not so painful. I tried to replace imports with lazy imports as a workaround but it's still 10 time worse in compare with current version using old generator.

My tests:

1. kubectl, local cluster (as a reference), get pods (9) from all namespaces - 0.12 sec
2. this library, current version, example1.py to list pods - 0.50 sec
3. OpenAPIGenerator v 7.0.1, example1.py - 17.26 sec
4. OpenAPIGenerator like above with some lazy imports (for models, apis) - 5.22 sec

The problem is known and may be resolved in the future: pydantic/pydantic#6748

For now I' m going to switch to the version just before Pydantic. It helps to resolve some pending issues, and this issue is for tracking ideas or attempts for future upgrades.

[multani]

@tomplus maybe it would be worth having a generator option to not load all the models from the __init__.py files (both client/__init__.py and client/models/__init__.py)

Typically, I'm using only a very small subset of all the available models, but importing kubernetes_asyncio always loads them all.

If we could reduce this default list to something smaller, and have to explicitly import the extra classes you need, I think it could help a lot.

[tomplus]

I did such tests using lazy imports (models were loaded only when needed) and it worked 3x faster but still 10x slower than the current version (see timings above). The problem is that each model requires other models and these models required others and still you need to load many of them...

[rad-pat]

Is this possible now they claim to have improved the speed? pydantic/pydantic#6748 (comment)

[jaseemabid]

I noticed that the generated client still depends on six, which should be unnecessary these days. Is that also blocked on using the new generator? Now that the Pydantic slow down is fixed, could we give it another try?

[tomplus]

Unfortunately it's better but still bad :/

Now I compare only import time:

a) current version

$ time PYTHONPATH=. python -m kubernetes_asyncio.client.api_client

real 0m0,431s
user 0m0,367s
sys  0m0,064s

b) generated by https://github.com/OpenAPITools/openapi-generator/releases/tag/v7.16.0

with libraries:

pydantic==2.12.3
pydantic-core==2.41.4

results

real 0m12,132s
user 0m11,860s
sys  0m0,265s

So it's 12sec vs 0.5 sec. Lazy imports will help but it's not a game changer here.