[Bug] fix RayActorOptionSpec.items.spec.serveConfig.deployments.rayActorOptions.memory int32 data type

[TsykunovDmitriy]

Search before asking

• I searched the issues and found no similar issues.

KubeRay Component

ray-operator

What happened + What you expected to happen

1. Amount of serve deployment's memory expected in bytes. Data type for memory expected in int32 for RayActorOptionSpec.items.spec.serveConfig.deployments.rayActorOptions.memory (link to source code). In my case, I cannot specify amount of memory equal to 3.5Gb (logs from kuberay-operator in 3 section).
2. Expected behavior - kuberay-operator supports uint64 data type (or int64).
3. Logs from kuberay-operator:

W0628 20:13:46.926544       1 reflector.go:324] pkg/mod/k8s.io/[email protected]/tools/cache/reflector.go:167: failed to list *v1alpha1.RayService: json: cannot unmarshal number 3758096384 into Go struct field RayActorOptionSpec.items.spec.serveConfig.deployments.rayActorOptions.memory of type int32
E0628 20:13:46.926578       1 reflector.go:138] pkg/mod/k8s.io/[email protected]/tools/cache/reflector.go:167: Failed to watch *v1alpha1.RayService: failed to list *v1alpha1.RayService: json: cannot unmarshal number 3758096384 into Go struct field RayActorOptionSpec.items.spec.serveConfig.deployments.rayActorOptions.memory of type int32

Reproduction script

Just create random serve deployment with memory requests over than 2Gb. In my case next deployment raise error from above log.

from ray import serve

@serve.deployment(
    ray_actor_options={"memory": 3.5 * 1024 * 1024 * 1024}
)
class DummyApp:
    pass

app = DummyApp.bind()

Anything else

Thread in ray slack.
Ping: @kevin85421

Are you willing to submit a PR?

• Yes I am willing to submit a PR!

[kevin85421]

I found something interesting: the memory field in the Ray Serve schema is specified as type float, and it is measured in bytes. Let me figure this out.

https://sourcegraph.com/github.com/ray-project/ray@ac2230f/-/blob/python/ray/serve/schema.py?L43-L49

[kevin85421]

[Update]

I found something interesting: the memory field in the Ray Serve schema is specified as type float, and it is measured in bytes.

The memory field is used to specify logical memory rather than physical memory. That is why it is of type float. It is very confusing.

I have already opened a PR #1220 to fix this issue. However, if you are comfortable with the nightly KubeRay, I will strongly recommend you use serveConfigV2 rather than serveConfig. serveConfig will be deprecated in the future. You can specify more than 2GB in serveConfigV2 without #1220.

[TsykunovDmitriy]

With serveConfigV2 it works as expected. Thanks.