Ray Serve

Ray Serve now lets each deployment return a 429 (optionally with a Retry-After header) instead of 503 when max_queued_requests is exceeded via a new BackpressureConfig, and introduces a TracingConfig model for serve.start() or config files that wires tracing through proxy and replica paths with fail-fast setup errors. Teams must migrate from environment-variable-only tracing setup to the new TracingConfig model, since tracing configuration no longer relies solely on env vars.

Matters for: Teams running Ray Serve behind load balancers or retry/SLO policies that need to distinguish deliberate load shedding from deployment failures, and teams using gang-scheduled autoscaling deployments that want to scale to zero replicas.