[BUG]: dynamographdeployment operator didn't handle multinode spec update operation

[Monokaix]

Describe the Bug

Current behavior:
When set multinode count=2 and enable lws, the dynamographdeployment operator will create a lws with 2 workers, and when didn't enable multinode, the operator will just create a deployment.

But if we deploy a backed with aggregated mode without setting multinode filed, and then update the dgd cr with setting multinode count=2, the previous created deployment will not be deleted, and the lws and deployment co-exist.

Similarly, if we deploy a backend with setting multinode=2 and then remove this field, the lws will not be removed either.

I'd like to know if update the multinode spec is a not allowed behavior, if true we should add the validation here and if not we should handle this correctly.

Steps to Reproduce

1. deploy a vllm backend with agg mode using this example
2. edit the dgd cr and set multinode=2
3. use kubectl and find that both deployment and lws exist.

Expected Behavior

The deployment should be deleted when set multinode field

Actual Behavior

Both deployment and lws exist and the deployment is not deleted.

Environment

This issue is environment-independent.

Additional Context

No response

Screenshots

No response

[julienmancuso]

Hello, yes this shouldn't be allowed and blocked by the operator during validation.
This should be done in the validation webhook because during reconciling phase, the operator can't determine the previous configuration.

Workaround while this is being implemented : delete your current DGD and re-create it

[Monokaix]

Hello, yes this shouldn't be allowed and blocked by the operator during validation. This should be done in the validation webhook because during reconciling phase, the operator can't determine the previous configuration.

Workaround while this is being implemented : delete your current DGD and re-create it

Thanks for your reply!
But seems a little aggressive to delete the whole cr, because it will also delete the frontend service, my though is that just delete the previous deployment or lws is enough?

[julienmancuso]

Yes you can also manually delete the deployments