[Feature] Remote call timeout required.

[MissiontoMars]

Search before asking

• I had searched in the issues and found no similar feature requirement.

Description

Why the feature needed?

Feature required from our users. The current remote call design lacks a timeout mechanism, actor.foo.remote() will be executed in the end. The user may expect the remote call to be completed in a certain period of time, if not, then give up. If the producer actor send lots of calls in a short time, some calls which destined to be timeout will also be executed, that's unnecessary.

API

Set timeout_ms option when do remote call. e.g.:

@ray.remote
    class Foo:
        def method(self):
            return 1
    f = Foo.remote()
    f.method.options(timeout_ms=500).remote()

When the caller actor send out the task, put the current timestamp in taskspec, and on the callee actor side, check the timestamp with current system time when prepare to execute the task, if timeout, give up.

Other considerations

1. The user can do the similar things by themself. Add an extra timestamp parameter in mthod, and do not execute the bussiness code when timeout. But I think it would be more convenient to support this option in Ray.
2. The time synchronization problem in distributed systems, makes the timeout not 100% accurate. But in most cases it's not a problem.

Use case

No response

Related issues

No response

Are you willing to submit a PR?

• Yes I am willing to submit a PR!

[jovany-wang]

Nice proposal.

1. Can you add the Java API or usage as well?
2. Is the implementation that adding a timestamp to the task execution information?

[mwtian]

What about child task and method calls? Do timeout / deadline propagate implicitly or explicitly?

[ericl]

@mwtian , child tasks are automatically killed when the parent dies, so there's nothing that should need to happen for that to work.

[pcmoritz]

+1 on the proposal, but it should be timeout and a float that is measured in seconds (instead of timeout_ms as an integer in milliseconds). That's to be consistent with the other timeouts in Ray and also Python timeouts.

[kfstorm]

@ericl Does the task canceling API apply to actor tasks?