Running on ECS Fargate

I’m trying to get Ray running on ECS Fargate. I have implemented a NodeProvider which launches ECS tasks with the latest ray docker images. The NodeProvider seems to expect bare iron however and it tries to ssh into the container and install docker and perform other setup. How can I skip all this and just tell Ray that I started a worker and it is up and running? Is there a different interface I should be implementing? I was using the KubernetesNodeProvider as a template but that seems to also ssh into the pods and perform setup. Thanks.


Have you managed to get this to work? I’m also trying to use Fargate, not sure how to implement NodeProvider solving issues you described

Hi @shoelesself and @dirtyValera the story for autoscaling with cluster managers is indeed a bit underdeveloped at the moment.

We (Ray maintainers) will think about how to improve the situation.

Hi @Dmitri, have the capabilities of Ray changed in the last 6 months to make it play nicer with cluster managers? (e.g. running a custom hook every time the utilization passes a threshold, like Dask does). At the moment, it seems difficult to auto-scale the cluster if you do not have direct access to the underlying Kubernetes.

The situation is pretty much the same. We haven’t been able to prioritize improving the interface.

I have very recently been asked about Fargate again, so it would be nice to think about how that might work. At the moment, we’re prioritizing improving support for K8s and VMs.

1 Like

Tracking this here:

We’ll get to it one day…