I set up a Ray cluster on Kubernetes and hope that multiple users can use it at the same time. However, during the test, I found a problem:
When a user uses
joblib, it seems that they always occupy all the resources of the cluster (such as CPU cores) as much as possible;
This results in another user’s task cannot be executed because of insufficient resources (for example, the number of CPU cores).
ray.util.connect() method to connect to the Ray cluster, such as the following code:
import ray ray.util.connect('<ray_cluster_ip>:<port>') import modin.pandas as pd """ User's code here """
Is it possible for Ray to provide such a function: when the client connects to the ray cluster, it can set the resource limit for the client to use. That is, from the perspective of the client, only part of the resources of the cluster is available. In this way, when multiple users use cluster resources, they can be separated.