Why ray serve need KubeRay to use GCS recover feature?

Ray Libraries (Data, Train, Tune, Serve) Ray Serve
1

Hello!
I was confused when I read this in document.

image
image1446×596 64.2 KB

If Ray can use Redis as persistence storage, why it need to depends on KubeRay?

Another question is when a head node is dead, can I use a new machine as a head node of the cluster, and recover from the Redis?

2

If Ray can use Redis as persistence storage, why it need to depends on KubeRay?

Ray needs an external mechanism to restart the head node if it dies. KubeRay monitors the Ray cluster’s status and restarts the head node when you enable GCS fault tolerance.

Another question is when a head node is dead, can I use a new machine as a head node of the cluster, and recover from the Redis?

Yes, I believe so.