github avandeputte/haproxy-manager v2.3.2
2.3.2: tcp:// services keep their connections

one hour ago

2.3.2: tcp:// services keep their connections for as long as they say

A fix for tcp:// services, from #35. Nothing changes for HTTP services.

What was wrong

A tcp:// service with a Server timeout of, say, 30 minutes still lost its connections after 50 seconds. HAProxy holds each half of a TCP connection to its own timeout, and only the server half got the service's value; the listener kept the 50 s client timeout that stands for everything. An idle database session was cut after 50 s, and so was a query that ran longer than that. Every tcp:// recipe sets a long server timeout and was affected.

What changed

The Server timeout of a tcp:// service is now also written as the pool's timeout tunnel, which replaces both halves:

backend be_pg
    mode tcp
    timeout server 30m
    timeout tunnel 30m

It takes effect at the next Apply after the update. Connections on tcp:// services then live as long as the Server timeout says, which is longer than before.

  • A tcp:// service without a Server timeout of its own is unchanged: it keeps the client and server timeouts under Advanced · HAProxy › Settings.
  • A pool with its own timeout tunnel in Extra directives keeps that one.
  • A timeout client added to a listener's Extra directives as a workaround is no longer needed and can be removed.

The guide's Tuning page describes it.

Tested

On the container image (HAProxy 3.4), with a 30 minute Server timeout: an idle connection and a connection waiting on a silent server were both cut at 50.0 s on 2.3.1, and were still open at 70 s with the fix. A 10 s Server timeout still cuts at 10 s. The two configuration lines were also checked on HAProxy 3.0. Not tried: a package install under systemd, which generates the same configuration.

Don't miss a new haproxy-manager release

NewReleases is sending notifications on new releases.