[RLlib] MultiAgentEnv would NOT call env.close() on a failed sub-env (if restart_failed_sub_environments is True).

[sven1977]

MultiAgentEnv would NOT call env.close() on a failed sub-env (if restart_failed_sub_environments is True).

Note: Single-agent works ok, because it's using a different VectorEnv base class.

Note: New stack EnvRunners are also already behaving correctly (making the Env.close() call upon failure).

Why are these changes needed?

Related issue number

Checks

• I've signed off every commit(by using the -s flag, i.e., git commit -s) in this PR.
• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
  • I've added any new APIs to the API Reference. For example, if I added a
    method in Tune, I've added it in doc/source/tune/api/ under the
    corresponding .rst file.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[simonsays1980]

LGTM. Only two questions:

1. Why do we deprecate via @OldStack the BaseEnv when the multi-agent wrapper still needs it?
2. We should probably expect a certain interface to be implemented such that each env we wrap here has a close method.

[sven1977]

LGTM. Only two questions:

1. Why do we deprecate via @OldStack the BaseEnv when the multi-agent wrapper still needs it?
2. We should probably expect a certain interface to be implemented such that each env we wrap here has a close method.

The MultiAgentWrapper is also only there b/c of BaseEnv.
Note also that OldAPIStack does not deprecate things yet. It merely marks that these classes belong to a set of APIs that we will no longer advance and develop and - at some point in the future - will be deprecated.