[Serve] Fix gang scheduling PG leak when deployment actors are starting

[harshit-anyscale]

Summary

Fix a placement group leak in gang scheduling when deployment actors are configured.

Root cause

When a deployment uses both gang scheduling and deployment actors, _reserve_gang_placement_groups() creates placement groups every control loop tick (~100ms) while deployment actors are starting. But scale_deployment_replicas() defers replica creation until actors are ready, leaving the PGs orphaned. After ~2 seconds, all cluster resources are consumed by phantom PGs and replicas can never be scheduled — deadlock.

The sequence on every tick while deployment actors are starting:

1. _reserve_gang_placement_groups(): replica delta > 0 → creates PG (reserves 2 CPUs)
2. scale_deployment_replicas(): deployment actors not ready → returns empty upscale → PG unused
3. Next tick: creates another PG (delta still > 0), still empty upscale
4. After ~18 ticks (~2s): all 36 CPUs consumed by orphaned PGs, zero resources left

Fix

Skip PG reservation in _reserve_gang_placement_groups() when deployment actors are configured but not yet ready, matching the precondition already in scale_deployment_replicas().

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces a serve_instance fixture in test_deployment_actors.py to restart the Serve controller between tests, preventing performance degradation caused by accumulated state. Feedback suggests replacing delete_all_apps() with serve.shutdown() in the fixture teardown to more reliably ensure a clean state if the controller becomes unresponsive.

[jeffreywang88]

signing off for the gang scheduling pieces