Default actor resources

What is the problem?

Ray documentation clearly indicates that by default, when not specified explicitly, ray actors will use 1 cpu resource (https://docs.ray.io/en/master/actors.html#resources-with-actors). However it seems like the default is actually 0.

Ray version and other system information (Python version, TensorFlow version, OS): I saw this behavior on Ray 0.8.4, 1.0.0, 1.1.0, and 2.0.0dev.
I’m using Python 3.7.1 and 3.8.5 on Ubuntu 16.04.

Reproduction (REQUIRED)

import ray
ray.init(num_cpus=1)

@ray.remote
class foo:
    def f(self):
        return 1

a = foo.remote()
b = foo.remote()     # There should not be enough resources for this
print (ray.get(a.f.remote()))
print (ray.get(b.f.remote()))    # This should fail (block indefinitely)

Based on the documentation, I would expect that there would be a warning that an actor could not be scheduled and is pending , and the second ray.get will block forever. However, this is not the case – code terminates, showing both actors were constructed even though total cpu resources was limited to 1.

In contrast, with explicit num_cpus parameter, the code behaves as expected.

import ray
ray.init(num_cpus=1)

@ray.remote(num_cpus=1)
class foo:
    def f(self):
        return 1

a = foo.remote()
b = foo.remote()     # There should not be enough resources for this
print (ray.get(a.f.remote()))
print (ray.get(b.f.remote()))    # This should fail (block indefinitely)

• I have verified my script runs in a clean environment and reproduces the issue.
• I have verified the issue also occurs with the latest wheels.

[wuisawesome]

Yeah this is a little subtle. ray.remote and ray.remote(num_cpus=1) are not the same. ray.remote means the actor uses CPUs whenever it runs anything (the constructor, a function, etc), but it frees the CPU when it’s not using it.

ray.remote(num_cpus=1) means it just always keeps the CPU.

The underlying actor creation task always requires a CPU, but hopefully as an end user, you don’t need to think about that :p

[wuisawesome]

We can close the issue then right?