Specify port when using ray.init() to start new local instance

Ray Clusters
1

How severe does this issue affect your experience of using Ray?

  • High: It blocks me to complete my task.

In my service code, I am connecting to an existing ray cluster.

In local tests, I’m using ray.init() to start a local instance in a fixture and pass this as the head endpoint to the service code.

The problem is that using ray.init() to start a local instance opens a random port, so the endpoint always changes.

I do not see any options to specify the port for a new local instance.

Right now I am doing the following in a test fixture:

ray.init(local_mode=True, ignore_reinit_error=True, num_cpus=1)

But this fails because the service ray.init() tries to connect on a ray endpoint that is specified in the environment, does not find the local instance because the local port doesn’t match the environment, and times out.

So I want to do something like this in a test fixture:

RAY_PORT=10001 # Specified in the environment
ray.init(local_mode=True, ignore_reinit_error=True, num_cpus=1, port=RAY_PORT)

The goal is to then be able to connect on f"ray://127.0.0.1:{RAY_PORT}" using ray.init() in the code.

Thanks!

2

Hello! What version of Ray are you using? local_mode seems to be deprecated as of the latest version of Ray. local_mode also doesn’t really set up a Ray cluster (like you may have noticed) since it is more of a ephemeral process used for debugging.

Have you tried using the regular ray.init (docs here: ray.init — Ray 2.42.1) and setting variable dashboard_port to a specific port?

Here’s a guide on how to customize the dashboard port using ray.init() if it helps:

https://docs.ray.io/en/latest/cluster/configure-manage-dashboard.html

3

Thanks, @christina!

I’m using ray 2.34.0

The dashboard port isn’t the same as the head node port, which is what needs to be passed to ray.init(address=xxxx) in order to connect to an existing cluster. Setting the dashboard port doesn’t impact the issue I am describing.

4

Oh, I see what you mean! Ok, so this is a way to start Ray with a port via the CLI (ray start --head --port=6379) Starting Ray — Ray 2.42.1

Having a stable port for every launch might help with the issues you’re facing. But since you’re doing it in a testing environment I think doing it in Python will be better than the CLI. The general idea is the same as the one above. I think the equivalent in Python code will be ray.init(address="127.0.0.1:6379") , or something like this in your test suite:

  ray.init(
      address=127.0.0.1:6379,      ...
  )

I think either one will work in setting the port, lmk if this helps in your tests.

5

Thanks, @christina, but this won’t work because if an address is specified, ray.init() will try to connect to it and not create an instance.

image
image2822×324 59.3 KB

According to the docs, the only way to have ray.init() create a new local instance is not to specify an address.

And yes, I cannot use the cli because the init needs to happen in the test only.

6

FYI I resolved this for now by using the address="auto" argument in ray.init()

# project/utils.py
def is_ray_local():
    """Logic to determine whether you are running locally or in a deployment"""
    return True

# project/service.py
import ray

from project.utils import is_ray_local


RAY_LOCAL_FLAG = is_ray_local()
RAY_CLUSTER_ENDPOINT_ACTUAL = "ray://1.1.1.1:11111"

def get_ray_enpoint()
    if RAY_LOCAL_FLAG:
        return "auto"
    return RAY_CLUSTER_ENDPOINT_ACTUAL

def foo_service():
    ray.init(address=get_ray_enpoint())
    # Do stuff
    return "Success"

# tests/test.py
import ray

from project.service import foo_service


def test_foo_service():
    ray.init()

    status = foo_service()

    assert status == "Success"
7

Ok! Thanks for the update :slight_smile: I read the docs for a bit and here’s just a small update with what I found out so far. I think your workaround works well, too. ^^ You can also set the environment variable RAY_ADDRESS and then ray.init() will automatically use the environment variable that you set previously automatically. I think this should work even if you are running test suites. Apologies, I wasn’t much help, but please feel free to reach out anytime!