Add core API and backend service design doc

[Jeffwan]

Why are these changes needed?

Per Ali's request in #93, It's better to write a one page summary about the proposed change and get it approved before any implementation.

Related issue number

#53 #54

/cc @akanso @chenk008 @caitengwei @chaomengyuan

Checks

• I've made sure the tests are passing.
• Testing Strategy
  • Unit tests
  • Manual tests
  • This PR is not tested :(

[Jeffwan]

I think ray-operator should not be aware of the imageTemplate. it just need image to build the cluster. ImageTemplate corresponding services can be used to build images separately.

1. User who does not have docker experience can build images using this service.
2. User can query the image result built by the service
3. User can use that image result to build ray cluster.

[Jeffwan]

/cc @akanso please have a check

[akanso]

The functionality of exposing a simpler interface to the users other than the K8s interface is very valuable to none K8s users.
GRPC is a generic interface, many clients in GoLang, C#, Python etc. can use it.

It is not clear from the design Doc if the operator code itself is extended to support this feature?

I think an alternative design, is to run the GPRC server as an independent container (e.g. side-car container), that is capable of processing the GRPC requests and create RayCluster Custom Resources that the Ray Operator can read and process.

client --> GRPC Server --> [created Custom Resources] <-- Ray Operator (reads CR and accordingly performs CRUD )

[Jeffwan]

@akanso

It is not clear from the design Doc if the operator code itself is extended to support this feature?

I see. I think the gap is I talked too much api etc instead of operator interactions and deployment topology in the doc.

I think an alternative design, is to run the GPRC server as an independent container (e.g. side-car container), that is capable of processing the GRPC requests and create RayCluster Custom Resources that the Ray Operator can read and process.
client --> GRPC Server --> [created Custom Resources] <-- Ray Operator (reads CR and accordingly performs CRUD )

exactly. It would be a separate pod and the path is like what you described. Let me add a section for it

[Jeffwan]

@akanso I add a short summary to talk about deployment and interactions.

[Jeffwan]

Ali reviewed the docs and overall looks good to him as well. We can merge the PR.