Connection Draining Patterns

Graceful LB removal.

Overview

Connection draining is what makes deploys zero-downtime. Backends remove themselves from the load balancer cleanly, in-flight requests complete, then the process exits. Without it, every deploy drops a slice of requests and customers see the difference. With it, deploys become invisible.

• Graceful LB removal. In-flight requests complete before the backend stops accepting new ones. Zero-error deploys.
• Drain timeout. How long to wait for in-flight to complete. Matches the longest expected request duration.
• Health check status. Backend reports unhealthy to start the drain. Load balancers stop sending new traffic; existing traffic completes.
• SIGTERM handling plus long-connection close. Application catches SIGTERM and finishes gracefully; WebSocket and gRPC streams need explicit close handshakes.

The approach

Three habits make connection draining real rather than theoretical: health-check-driven drain, drain timeout tuned to actual request profile, and SIGTERM handlers that stop accepting new work while finishing in-flight requests.

• Health check drives drain. Backend reports unhealthy before process exit. Load balancer stops sending new traffic immediately.
• Drain timeout tuned to longest expected request. Avoids the cutoff that drops in-flight requests.
• SIGTERM handler. Application stops accepting new connections, finishes existing ones, then exits. Standard Kubernetes pattern.
• Long-connection explicit close plus documented policy. WebSocket and gRPC explicit close handshakes; per-service the drain timeout documented.

Why this compounds

Each clean deploy preserves user experience that bad deploys would have damaged. The team’s deploy hygiene deepens; new services inherit the drain conventions instead of recreating them through customer-impact incidents.

• User experience preserved. Zero-error deploys keep trust intact across hundreds of releases.
• Release safety improves. Clean deploys reduce incident rate. Velocity rises.
• Operational maturity. Drain patterns scale to every service. The estate operates the same way.
• Year-one investment, year-two habit. First service takes care. By the third, drain handlers are part of the deploy template.