Cloud Setup and Auto-Scaling

The ray create_or_update command starts an AWS Ray cluster from your personal computer. Once the cluster is up, you can then SSH into it to run Ray programs.

Quick start

First, install boto (pip install boto3) and configure your AWS credentials in ~/.aws/credentials, as described in the boto docs.

Then you’re ready to go. The provided ray/python/ray/autoscaler/aws/example-full.yaml cluster config file will create a small cluster with a m5.large head node (on-demand) configured to autoscale up to two m5.large spot workers.

Try it out by running these commands from your personal computer. Once the cluster is started, you can then SSH into the head node, source activate tensorflow_p36, and then run Ray programs with ray.init( + ":6379").

# Create or update the cluster. When the command finishes, it will print
# out the command that can be used to SSH into the cluster head node.
$ ray create_or_update ray/python/ray/autoscaler/aws/example-full.yaml

# Reconfigure autoscaling behavior without interrupting running jobs
$ ray create_or_update ray/python/ray/autoscaler/aws/example-full.yaml \
    --max-workers=N --no-restart

# Teardown the cluster
$ ray teardown ray/python/ray/autoscaler/aws/example-full.yaml

To run connect to applications running on the cluster (e.g. Jupyter notebook) using a web browser, you can forward the port to your local machine using SSH:

$ ssh -L 8899:localhost:8899 -i <key> <user>@<addr> 'source ~/anaconda3/bin/activate tensorflow_p36 && jupyter notebook --port=8899'

Updating your cluster

When you run ray create_or_update with an existing cluster, the command checks if the local configuration differs from the applied configuration of the cluster. This includes any changes to synced files specified in the file_mounts section of the config. If so, the new files and config will be uploaded to the cluster. Following that, Ray services will be restarted.

You can also run ray create_or_update to restart a cluster if it seems to be in a bad state (this will restart all Ray services even if there are no config changes).

If you don’t want the update to restart services (e.g. because the changes don’t require a restart), pass --no-restart to the update call.


By default, the nodes will be launched into their own security group, with traffic allowed only between nodes in the same group. A new SSH key will also be created and saved to your local machine for access to the cluster.


Ray clusters come with a load-based auto-scaler. When cluster resource usage exceeds a configurable threshold (80% by default), new nodes will be launched up the specified max_workers limit. When nodes are idle for more than a timeout, they will be removed, down to the min_workers limit. The head node is never removed.

The default idle timeout is 5 minutes. This is to prevent excessive node churn which could impact performance and increase costs (in AWS there is a minimum billing charge of 1 minute per instance, after which usage is billed by the second).

Monitoring cluster status

You can monitor cluster usage and auto-scaling status by tailing the autoscaling logs in /tmp/raylogs/monitor-*.

The Ray autoscaler also reports per-node status in the form of instance tags. In the AWS console, you can click on a Node, go the the “Tags” pane, and add the ray:NodeStatus tag as a column. This lets you see per-node statuses at a glance:


Customizing cluster setup

You are encouraged to copy the example YAML file and modify it to your needs. This may include adding additional setup commands to install libraries or sync local data files.


After you have customized the nodes, it is also a good idea to create a new machine image (AMI) and use that in the config file. This reduces worker setup time, improving the efficiency of auto-scaling.

The setup commands you use should ideally be idempotent, that is, can be run more than once. This allows Ray to update nodes after they have been created. You can usually make commands idempotent with small modifications, e.g. git clone foo can be rewritten as test -e foo || git clone foo which checks if the repo is already cloned first.

Most of the example YAML file is optional. Here is a reference minimal YAML file, and you can find the defaults for optional fields in this YAML file.

Syncing git branches

A common use case is syncing a particular local git branch to all workers of the cluster. However, if you just put a git checkout <branch> in the setup commands, the autoscaler won’t know when to rerun the command to pull in updates. There is a nice workaround for this by including the git SHA in the input (the hash of the file will change if the branch is updated):

file_mounts: {
    "/tmp/current_branch_sha": "/path/to/local/repo/.git/refs/heads/<YOUR_BRANCH_NAME>",

    - test -e <REPO_NAME> || git clone<REPO_ORG>/<REPO_NAME>.git
    - cd <REPO_NAME> && git fetch && git checkout `cat /tmp/current_branch_sha`

This tells ray create_or_update to sync the current git branch SHA from your personal computer to a temporary file on the cluster (assuming you’ve pushed the branch head already). Then, the setup commands read that file to figure out which SHA they should checkout on the nodes. Note that each command runs in its own session. The final workflow to update the cluster then becomes just this:

  1. Make local changes to a git branch
  2. Commit the changes with git commit and git push
  3. Update files on your Ray cluster with ray create_or_update

Common cluster configurations

The example-full.yaml configuration is enough to get started with Ray, but for more compute intensive workloads you will want to change the instance types to e.g. use GPU or larger compute instance by editing the yaml file. Here are a few common configurations:

GPU single node: use Ray on a single large GPU instance.

max_workers: 0
    InstanceType: p2.8xlarge

Docker: Specify docker image. This executes all commands on all nodes in the docker container, and opens all the necessary ports to support the Ray cluster. This currently does not have GPU support.

    image: tensorflow/tensorflow:1.5.0-py3
    container_name: ray_docker

Mixed GPU and CPU nodes: for RL applications that require proportionally more CPU than GPU resources, you can use additional CPU workers with a GPU head node.

max_workers: 10
    InstanceType: p2.8xlarge
    InstanceType: m4.16xlarge

Autoscaling CPU cluster: use a small head node and have Ray auto-scale workers as needed. This can be a cost-efficient configuration for clusters with bursty workloads. You can also request spot workers for additional cost savings.

min_workers: 0
max_workers: 10
    InstanceType: m4.large
        MarketType: spot
    InstanceType: m4.16xlarge

Autoscaling GPU cluster: similar to the autoscaling CPU cluster, but with GPU worker nodes instead.

min_workers: 0
max_workers: 10
    InstanceType: m4.large
        MarketType: spot
    InstanceType: p2.xlarge

External Node Provider

Ray also supports external node providers (check implementation). You can specify the external node provider using the yaml config:

    type: external
    module: mypackage.myclass

The module needs to be in the format package.provider_class or package.sub_package.provider_class.

Additional Cloud providers

To use Ray autoscaling on other Cloud providers or cluster management systems, you can implement the NodeProvider interface (~100 LOC) and register it in Contributions are welcome!