[Data] Release dead actors from the actor pool so they can be replaced

[lonexreb]

Why are these changes needed?

With a fixed-size actor pool and max_errored_blocks = -1, an actor that dies mid-pipeline (e.g. sys.exit(0) inside the UDF) makes the pipeline hang silently forever (repro from #62746, verified on master: hangs indefinitely; with this PR it completes).

Root cause, as diagnosed in the issue:

• refresh_actor_state detects the actor as DEAD and _update_rank removes it from the scheduling heaps, so no new tasks are dispatched to it — but the entry stays in _running_actors.
• The pool's current_size() therefore never drops, so for a fixed-size pool (min_size == max_size) the autoscaler never creates a replacement (default_actor_autoscaler only upscales when current_size() < min_size()).
• With max_errored_blocks = -1, the failed block is ignored instead of aborting, leaving 0 schedulable actors, a non-empty queue, and no exception.

The fix:

• refresh_actor_state now releases definitively-DEAD actors from the pool (with a warning log); the autoscaler then scales the pool back up to its min size. Actors whose local state is None (unknown, possibly transient) are left in place, as before.
• on_task_completed tolerates actors already released with tasks in flight — _release_running_actor has already reconciled the pool's counters for them.

Related issue number

Fixes #62746

Checks

• I've signed off every commit (DCO).
• Formatting with the repo-pinned black==22.10.0.
• Testing strategy (run locally against master via Ray nightly wheel + setup-dev.py symlinks):
  • Issue's repro script (bounded by a watchdog alarm): hangs on master, completes with this PR — dead actors are released ("actor ... is dead; releasing it from the actor pool so it can be replaced") and replaced, and materialize() finishes with the errored blocks dropped.
  • Added test_dead_actor_released_and_replaced_e2e: size-1 pool, UDF exits once via sys.exit(0), asserts the pipeline completes. Passes in ~7s (pre-fix it would hang into the 180s pytest timeout).
  • pytest python/ray/data/tests/test_actor_pool_map_operator.py → 54 passed.

Notes for reviewers / AI-assistance disclosure

• This PR was prepared with AI assistance; I have reviewed every changed line and run the tests above locally.
• Duplicate-work check: no open PR references [Data] Dead Actor occupies actor pool slots which severely effects resource usage #62746 and the issue has no comments. The issue author checked "willing to submit a PR" in April 2026 but none has appeared since; happy to defer if one is in progress.

[gemini-code-assist]

Code Review

This pull request ensures that dead actors are properly released from the actor pool and replaced by the autoscaler, preventing pipelines from silently stalling. It updates refresh_actor_state to track and release dead actors, handles already-released actors in on_task_completed, and adds an end-to-end test to verify this behavior. There are no review comments, so I have no feedback to provide.

[github-actions]

This pull request has been automatically marked as stale because it has not had
any activity for 14 days. It will be closed in another 14 days if no further activity occurs.
Thank you for your contributions.

You can always ask for help on our discussion forum or Ray's public slack channel.

If you'd like to keep this open, just leave any comment, and the stale label will be removed.

[lonexreb]

Not stale — rebased onto current master; the regression test (test_dead_actor_released_and_replaced_e2e) passes locally in ~7s. Ready for review whenever a maintainer has bandwidth.