Hey folks. I have a Ray Cluster running on GCP and...
# ray-integration
p
Hey folks. I have a Ray Cluster running on GCP and am relatively new to MLOps -- from reading this blog post https://flyte.org/blog/ray-and-flyte I did not fully grasp the benefits of running Ray inside Flyte -- can some users of both explain the benefits to me?
k
i think the benefits are ephemeral cluster • Automatic setup and teardown cluster • Cluster has all your python dependencies installed and is reproducible • you can create one cluster per execution / per user and they can be independent
But, you do not have to use Flyte and KubeRay clusters
we infact will be working with Anyscale and others to add Flyte Agent support to send workloads to existing kubernetes clusters
how do you manage the lifecycle of the clusters across multiple users?
p
My usecase is maybe a bit different, there are no users other than my service, which launches predefined jobs on the cluster.
So Ray itself is never exposed to the users.
k
and you want the cluster to be up all the time?
p
Yeah, ideally:)
k
if so, I think adding a Flyte Agent would be easy or run Ray in client mode
the Flyte Ray job is for emphemeral, isolated and potentially "short" lived clusters
p
I see -- makes sense! Thanks for the advice
k
or simply run the ray head node in client mode in a flyte pod
k
cc @Yicheng Lu Ray agent
y
working on it