Lora controller doesn't clean up the stable endpoint ip

[Jeffwan]

🐛 Describe the bug

I notice the endpoint slice is not updated. I only have one pod but there're two ips. Considering we are doing upgrades, I think the model adapter controller doesn't handle this case well which brings the technical issues.

Steps to Reproduce

No response

Expected behavior

No response

Environment

No response

[Jeffwan]

this can be reproduced if we have some rolling upgrade happening.

I determine to use this way to update the list instead of using append. When we start to support multiple replicas, we need some refactor anyway.

https://github.com/aibrix/aibrix/blob/3f6ce3ded76b9fed8d3837a60786800cf64d1d71/pkg/controller/modeladapter/modeladapter_controller.go#L862-L865

[Jeffwan]

There's a potential improvement, scheduler should pick up the new pod rather than old pod. Otherwise it will experience another round migration