TCP connections: the handshake, retransmission timers and keep-alives

What it is

RFC 9293 describes the three-way handshake: the client sends SYN, the server answers SYN-ACK and moves to SYN-RECEIVED, the client's ACK completes the connection. Every segment carrying data is acknowledged; if the acknowledgement does not arrive before the retransmission timeout (RTO) expires, the segment is sent again. RFC 6298 defines how the RTO is computed from measured round-trip times, with an initial value of 1 second before any measurement, and how it is doubled ("backed off") after each expiry. The RFC allows but does not require keep-alives: probes sent on an idle connection, with a default interval that must be no less than two hours.

Why it matters

The timers explain most "why did this hang for so long" reports. According to tcp(7), Linux retransmits an initial SYN tcp_syn_retries times (default 6, roughly 127 seconds) before a connect fails, and retransmits data in an established connection tcp_retries2 times (default 15, roughly 13 to 30 minutes) before giving up. Keep-alives are only sent when the socket has SO_KEEPALIVE set; the defaults are tcp_keepalive_time 7200 seconds, then 9 probes 75 seconds apart. A server whose peer vanished without a FIN or RST therefore keeps the connection, and its memory and file descriptor, for a long time.

How to apply

• Set application-level timeouts on connect, read and write; do not rely on TCP to detect dead peers quickly.
• Enable keep-alive on long-lived connections (database pools, message brokers, tunnels) and lower the idle time per socket (TCP_KEEPIDLE, TCP_KEEPINTVL, TCP_KEEPCNT in tcp(7)) rather than system-wide.
• Use TCP_USER_TIMEOUT where the library exposes it: it bounds how long unacknowledged data may remain outstanding before the connection is closed.
• When diagnosing, look at ss -tin for retransmit counters and RTO values, and at the connection state (SYN-SENT means no SYN-ACK ever came back).

Pitfalls

A completed handshake proves reachability at connection time only; a NAT or firewall dropping state later produces silence, not an error. Retransmission backoff is exponential, so a few consecutive losses turn into tens of seconds of stall. Keep-alive probes are ACK segments and are not reliably delivered, which is why the RFC forbids treating one unanswered probe as a dead connection.