Skip to main content

Transport Layer

The transport layer is the most common root cause of backend latency and reliability problems.

Why It Matters​

  • Connection setup and teardown directly affect tail latency.
  • Poor timeout/retry strategy amplifies outages.
  • Buffer and window tuning controls throughput on high-latency links.

TCP vs UDP​

TopicTCPUDP
ReliabilityOrdered and retransmittedBest-effort
ConnectionStatefulConnectionless
Typical UsageHTTP, databases, RPCDNS, streaming, QUIC transport

TCP Three-Way Handshake​

Handshake adds startup latency. Connection reuse is essential for high-QPS systems.

Flow Control and Congestion Control​

  • Flow control protects receiver buffers.
  • Congestion control protects the network path.

Monitor with:

ss -ti

Connection Lifecycle​

High TIME_WAIT counts are normal in short-connection workloads, but can still exhaust ephemeral ports.

Practical Tuning Areas​

  • Keep-alive and connection pool limits.
  • Connect/read/write timeout budget.
  • Retry policy with idempotency and backoff.
  • Kernel socket settings only after measurement.

Debugging Playbook​

# Socket states and queue sizes
ss -tan state established,time-wait

# Packet-level view
tcpdump -i any tcp port 443 -nn

Common Incidents​

Connection timeout​

  • Check route/firewall/listening port in order.
  • Verify timeout mismatch between caller and callee.

Connection reset​

  • Inspect RST packets and upstream idle timeout.
  • Verify keep-alive heartbeat and proxy settings.

Throughput collapse on long RTT​

  • Validate window scaling and receive buffers.
  • Compare congestion algorithm behavior by workload.