ControllerHealthMetrics#

pydantic model ray.serve.schema.ControllerHealthMetrics[source]#

Health metrics for the Ray Serve controller.

These metrics help diagnose controller performance issues, especially as cluster size increases.

PublicAPI (alpha): This API is in alpha and may change before becoming stable.

field application_state_update_duration_s: DurationStats | None = None#: Application state update duration statistics over a rolling window.

field controller_start_time: float = 0.0#: When the controller started (epoch seconds).

field deployment_state_update_duration_s: DurationStats | None = None#: Deployment state update duration statistics over a rolling window.

field event_loop_delay_s: float = 0.0#: Difference between actual and expected sleep duration. Positive values indicate an overloaded event loop.

field expected_sleep_duration_s: float = 0.0#: Expected sleep duration (CONTROL_LOOP_INTERVAL_S).

field handle_ingest_duration_ms: DurationStats | None = None#: Per-call processing time inside record_autoscaling_metrics_from_handle (rolling window, ms).

field handle_metrics_delay_ms: DurationStats | None = None#: Delay between when handle metrics are generated and when they reach the controller (rolling window, milliseconds).

field handle_reports_received: int = 0#: Total handle metric reports ingested since start.

field ingest_cpu_fraction: float = 0.0#: Fraction of one event-loop core consumed by the metrics ingestion path over the recent sampling window. Approaches 1.0 as ingestion monopolizes the single control-loop thread.

field ingest_reports_received: int = 0#: Total autoscaling metric reports ingested since start.

field last_control_loop_time: float = 0.0#: Time of the last control loop execution (epoch seconds).

field last_sleep_duration_s: float = 0.0#: Actual sleep duration of the last iteration.

field loop_duration_s: DurationStats | None = None#: Control loop duration statistics over a rolling window.

field loops_per_second: float = 0.0#: Control loop iterations per second.

field metrics_decompress_duration_ms: DurationStats | None = None#: Per-call decompress time for cloudpickle metric reports (rolling window, ms).

field node_update_duration_s: DurationStats | None = None#: Node update duration statistics over a rolling window.

field num_asyncio_tasks: int = 0#: Number of pending asyncio tasks.

field num_control_loops: int = 0#: Total number of control loops executed.

field process_memory_mb: float = 0.0#: Controller process memory usage in MB.

field proxy_state_update_duration_s: DurationStats | None = None#: Proxy state update duration statistics over a rolling window.

field replica_ingest_duration_ms: DurationStats | None = None#: Per-call processing time inside record_autoscaling_metrics_from_replica (rolling window, ms).

field replica_metrics_delay_ms: DurationStats | None = None#: Delay between when replica metrics are generated and when they reach the controller (rolling window, milliseconds).

field replica_reports_received: int = 0#: Total replica metric reports ingested since start.

field timestamp: float = 0.0#: When these metrics were collected (epoch seconds).

field uptime_s: float = 0.0#: Controller uptime in seconds.