Skip to content

Latest commit

 

History

History
378 lines (240 loc) · 10 KB

File metadata and controls

378 lines (240 loc) · 10 KB
title Graceful Shutdown

Graceful Shutdown

Graceful shutdown allows in-flight requests to complete before the gateway terminates, preventing request failures during deployments and restarts.


Overview

:material-power: Clean Termination

Allow existing requests to finish rather than abruptly closing connections.

:material-rocket-launch: Zero-Downtime Deployments

Deploy updates without causing client-visible errors.

:material-timer: Configurable Grace Period

Control how long to wait for in-flight requests.

:material-api: API Control

Trigger shutdown programmatically via HTTP API.


Why Graceful Shutdown?

Without graceful shutdown:

  • Abrupt termination: Active requests are immediately disconnected
  • Client errors: In-flight requests return connection errors
  • Data loss: Streaming responses may be truncated
  • Deployment failures: Rolling updates cause visible errors

With graceful shutdown:

  • Request completion: Active requests finish normally
  • No client errors: Users don't see deployment-related failures
  • Clean streaming: Streaming responses complete before shutdown
  • Smooth deployments: Zero-downtime rolling updates

How It Works

Graceful Shutdown Sequence

Shutdown Sequence

  1. Shutdown signal received (SIGTERM or SIGINT). The mesh-only /ha/shutdown API triggers a separate mesh-level broadcast path and is not part of this signal-driven sequence.
  2. Stop accepting new connectionsaxum_server's handle stops the TCP accept loop and marks the in-flight tracker as draining; new connections are refused at the socket level rather than receiving a 503 response. From this moment /readiness reports 503 (reason "draining") while /health and /liveness stay 200, so load balancers de-list the pod without restarting it.
  3. Drain in-flight requests — existing requests continue processing while the server waits on the in-flight tracker.
  4. Grace period timer starts — after --shutdown-grace-period-secs, the drain wait times out and the server forces shutdown with any remaining requests still in-flight.
  5. Clean exit — once all requests complete (or the grace period expires), background components (MCP orchestrator, etc.) are cleaned up and the process exits.

Configuration

smg \
  --worker-urls http://w1:8000 http://w2:8000 \
  --shutdown-grace-period-secs 180

Parameters

Parameter Default Description
--shutdown-grace-period-secs 180 (3 min) Time to wait for in-flight requests

Recommended Configurations

:material-lightning-bolt: Fast Shutdown

Quick termination for development.

smg --shutdown-grace-period-secs 10

Use when: Development, testing, quick restarts

:material-server-network: Production Standard

Balanced grace period for typical workloads.

smg --shutdown-grace-period-secs 180

Use when: Standard production deployments

:material-cog: Batch Processing

Long grace period for long-running requests.

smg --shutdown-grace-period-secs 600

Use when: Batch inference, long-running generations

:material-clock-fast: Critical Low-Latency

Minimal grace for latency-sensitive systems.

smg --shutdown-grace-period-secs 30

Use when: Very short requests, rapid scaling


Triggering Shutdown

Via Signal

# Find the SMG process
pgrep -f smg

# Send SIGTERM for graceful shutdown
kill -TERM <pid>

# Or SIGINT (Ctrl+C in terminal)
kill -INT <pid>

Via API

# Trigger graceful shutdown via HTTP (mesh mode only)
curl -X POST http://gateway:30000/ha/shutdown

The /ha/shutdown endpoint lives on the main gateway port (default 30000) and requires mesh mode (--mesh-* flags). Without mesh enabled the endpoint returns 503 Service Unavailable. The mesh handler broadcasts a LEAVING status to peer nodes and stops the mesh rate-limit task — it does not share the same in-flight drain path used by the signal handler.

Kubernetes Integration

Kubernetes sends SIGTERM by default when terminating pods. Configure terminationGracePeriodSeconds to match or exceed your SMG grace period:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: smg
spec:
  template:
    spec:
      terminationGracePeriodSeconds: 210  # SMG grace + buffer
      containers:
        - name: smg
          args:
            - --shutdown-grace-period-secs=180

!!! warning "Kubernetes timeout" Kubernetes will force-kill the pod after terminationGracePeriodSeconds. Set this higher than --shutdown-grace-period-secs to ensure SMG has time to complete its graceful shutdown.


Sizing the Grace Period

Consider these factors when setting the grace period:

Factor Impact on Grace Period
Average request duration Grace period should exceed typical request time
Longest expected request Batch jobs may need longer grace periods
Streaming responses Long streams need extended grace periods
Deployment frequency Frequent deployments may need shorter periods
Scaling responsiveness Autoscaling may need faster termination

Calculation Guidelines

grace_period = max(
    avg_request_duration * 3,
    p99_request_duration * 1.5,
    max_streaming_duration
)

Example: If your average request is 30s, p99 is 60s, and max streaming is 120s:

grace_period = max(90, 90, 120) = 120 seconds

Integration with Load Balancers

For zero-downtime deployments, coordinate with your load balancer:

Pre-Stop Hook (Kubernetes)

Remove the pod from the load balancer before shutdown:

spec:
  containers:
    - name: smg
      lifecycle:
        preStop:
          exec:
            command: ["/bin/sh", "-c", "sleep 5"]

The sleep allows the load balancer to stop sending new traffic before SMG begins its graceful shutdown.

Health Check Coordination

As soon as SMG receives the shutdown signal and begins draining, /readiness flips to 503 Service Unavailable with reason "draining", while /health and /liveness keep returning 200 OK throughout the drain. Kubernetes therefore removes the pod from Service endpoints (stopping new connections) without restarting it:

curl http://gateway:30000/health
# Returns 200 OK both during normal operation and throughout the drain

curl http://gateway:30000/readiness
# 200 while serving; 503 {"status":"not ready","reason":"draining"} once shutdown begins

/readiness also returns 503 when no healthy workers remain (or, in prefill/decode mode, when either side has no healthy worker), independent of the shutdown signal. The readiness decision is maintained event-driven from worker registry state and served from cached memory, so probes stay O(1) regardless of fleet size.

Dedicated Probe Port

Under heavy load the main listener's probe routes share the request runtime, so probe responses can lag behind request traffic. Pass the --health-check-port flag (Python: health_check_port) to additionally serve /liveness, /readiness, and /health on a dedicated plain-HTTP port, handled by a small isolated runtime on its own OS thread — probe latency then stays flat even when the request runtime is saturated, and the port keeps answering through the entire drain window:

spec:
  containers:
    - name: smg
      args: ["--health-check-port", "30001"]
      livenessProbe:
        httpGet: { path: /liveness, port: 30001 }
      readinessProbe:
        httpGet: { path: /readiness, port: 30001 }

When --health-check-port is unset no extra listener is started. The probe routes always remain available on the main port as well.


Monitoring

Shutdown Events

Watch logs for shutdown-related messages:

# Signal received
INFO Received Ctrl+C, starting graceful shutdown
# or
INFO Received terminate signal, starting graceful shutdown

# Gate — in-flight tracker is marked draining and the accept loop stops
INFO Beginning graceful shutdown: gating new connections in_flight=5

# Drain completes within the grace period
INFO All in-flight requests drained

# Or the grace period expires with requests still running
WARN Drain timed out, forcing shutdown with requests still in-flight remaining=2 timeout_secs=180

# Component teardown
INFO HTTP server stopped. Starting component cleanup...
INFO Cleanup complete. Process exiting.

Metrics During Shutdown

Metric Observation
smg_worker_requests_active Should decrease towards 0
smg_http_requests_total New requests should stop

Tuning Guidelines

Symptom Potential Adjustment
Requests failing during deployment Increase --shutdown-grace-period-secs
Slow scaling down Decrease --shutdown-grace-period-secs
Kubernetes force-killing pods Increase terminationGracePeriodSeconds
Streaming responses truncated Match grace period to max stream duration

What's Next?

:material-electric-switch: Circuit Breakers

Isolate failing workers to prevent cascade failures.

Circuit Breakers →

:material-heart-pulse: Health Checks

Proactive worker monitoring and failure detection.

Health Checks →

:material-traffic-light: Rate Limiting

Protect workers from overload with token bucket rate limiting.

Rate Limiting →