Files
kuberay/docs/guidance/rayStartParams.md
David XiaandGitHub b248d453e3 [chore][docs] enable Markdown lint rule MD013 (#3167)
* MD013: Line length

* Set max line length to 120 characters
* Ignore most files that violate the rule
* Fix some that only have a few violations
* Fix misspellings and formatting in `clients/python-client/README.md`

https://github.com/markdownlint/markdownlint/blob/main/docs/RULES.md

Signed-off-by: David Xia <[email protected]>
2025-03-07 15:49:11 -08:00

4.2 KiB

Default Ray Start Parameters for KubeRay

This document outlines the default settings for rayStartParams in KubeRay.

Options Exclusive to the Head Pod

  • --dashboard-host: Host for the dashboard server, either localhost (127.0.0.1) or 0.0.0.0. The latter setting exposes the Ray dashboard outside the Ray cluster, which is required when ingress is utilized for Ray cluster access. The default value for both Ray and KubeRay 0.5.0 is localhost. Please note that this will change for versions of KubeRay later than 0.5.0, where the default setting will be 0.0.0.0.

  • --no-monitor (Modification is not recommended):

    • Ray autoscaler supports various node providers such as AWS, GCP, Azure, and Kubernetes. However, the default autoscaler is not compatible with Kubernetes. Therefore, when KubeRay autoscaling is enabled (i.e. EnableInTreeAutoscaling is true), KubeRay disables the monitor process via setting --no-monitor to true and injects a sidecar container for KubeRay autoscaler. See PR #13505 for more details.
    • Please note that the monitor process serves not only for autoscaling but also for observability, such as Prometheus metrics. Considering this, it is reasonable to disable the Kubernetes-incompatible autoscaler regardless of the value of EnableInTreeAutoscaling. To achieve this, we can launch the monitor process without autoscaling functionality by setting the autoscaler to READONLY mode. If autoscaling-option is not set, the autoscaler will default to READONLY mode.
  • --port: Port for the GCS server. The port is set to 6379 by default. Please ensure that this value matches the gcs-server container port in Ray head container.

  • --redis-password: Redis password for an external Redis, necessary when fault tolerance is enabled. The default value is "" after Ray 2.3.0. See #929 for more details.

Options Exclusive to Worker Pods

  • --address: Address of the GCS server. Worker pods utilize this address to establish a connection with the Ray cluster. By default, this address takes the form <FQDN>:<GCS_PORT>. The GCS_PORT corresponds to the value set in the --port option. For more insights on Fully Qualified Domain Name (FQDN), refer to PR #938 and PR #951.

Options Applicable to Both Head and Worker Pods

  • --block: This option blocks the ray start command indefinitely. It will be automatically set by KubeRay. See PR #675 for more details. Modification is not recommended.

  • --memory: Amount of memory on this Ray node. Default is determined by Ray container resource limits. Modify Ray container resource limits instead of this option. See PR #170.

  • --metrics-export-port: Port for exposing Ray metrics through a Prometheus endpoint. The port is set to 8080 by default. Please ensure that this value matches the metrics container port if you need to customize it. See PR #954 and prometheus-grafana doc for more details.

  • --num-cpus: Number of logical CPUs on this Ray node. Default is determined by Ray container resource limits. Modify Ray container resource limits instead of this option. See PR #170. However, it is sometimes useful to override this autodetected value. For example, setting num-cpus:"0" for the Ray head pod will prevent Ray workloads with non-zero CPU requirements from being scheduled on the head.

  • --num-gpus: Number of GPUs on this Ray node. Default is determined by Ray container resource limits. Modify Ray container resource limits instead of this option. See PR #170.