Opscraft
The TCP handshake · Lesson 5 of 8 · 3 min read

Latency: handshakes, TLS and connection reuse

Counting round trips

Connection Round trips before the first request byte
TCP + HTTP (plain) 1
TCP + TLS 1.2 3
TCP + TLS 1.3 2
TLS 1.3 resumption with 0-RTT 1 (data in first TLS flight)
QUIC (HTTP/3) new connection 1 (transport + TLS combined)
QUIC 0-RTT resumption 0

With 150 ms RTT, a fresh TLS 1.2 connection costs ~450 ms before the request — which is why every hop should reuse connections:

  • HTTP keep-alive and HTTP/2 multiplexing between clients and load balancers.
  • Load balancers keep connections to targets open and reuse them across client requests (ALB does), avoiding a handshake per request.
  • Connection pools in apps for databases and downstream APIs.
  • TLS session resumption for repeat clients.
  • CDNs terminate TCP/TLS close to the user (short RTT), then reuse long-lived connections to the origin.
Enlarge diagramuserin ChennaiCDN edgenearbyoriginfar away20 ms: new TCP + TLS180 ms: kept open, reusedWithout the edge, every new visitor would pay 3 × 200 ms just to set up the connection.
Figure 1. Why CDNs and load balancers make sites faster: the expensive handshakes happen over a short distance, and the long-distance connections are kept open and reused.
Swipe sideways to see all of it, or tap to open it full screen.

Idle timeout mismatches: the 502 classic

A load balancer reuses an idle connection to a target at the exact moment the target's keep-alive timeout closes it → the request hits a closed connection → intermittent 502. Rule: the backend's keep-alive timeout must be longer than the load balancer's idle timeout (ALB default 60 s → set the app to e.g. 65–75 s; Node.js's default keepAliveTimeout of 5 s is a famous offender).