Description

Currently, Ray Serve's deployment scheduler only supports soft node label constraints when scheduling replicas.

In deployment_scheduler.py:596-600, NodeLabelSchedulingStrategy is always created with hard={}:

elif target_labels is not None:
    scheduling_strategy = NodeLabelSchedulingStrategy(
        hard={}, soft=target_labels
    )

While Ray Core's NodeLabelSchedulingStrategy fully supports both hard and soft constraints, Ray Serve only exposes the soft option.

Some workloads require strict node placement (e.g., compliance, hardware isolation, cost control) and should fail rather than fall back to unintended nodes.

Was it an intentional design decision to only support soft constraints in Serve (while Core supports both)? If so, I'd like to understand the reasoning.