go net/http server timeouts

don't blindly copy the timeouts...


go net/http server timeouts

don't blindly copy the timeouts...

net/http Server

Go's net/http.Server can be considered a production grade and suitable to exposing to the internet. There are, of course, several settings you may want to change, namely timeouts and crypto.

Filippo Valsorda has a few posts on that from a few years back: on timeouts short full, on everything.

Who your clients are

So I read through those, set some timeouts as suggested, and didn't think too much about it. Later, I moved the servers behind a reverse proxy, envoy, and everything worked fine. A few days later, I added health probes every 10 seconds.

error: unexpected status: 503 Service Unavailable body: upstream connect error or disconnect/reset before headers. reset reason: connection termination

Both envoy and my go server are on the same machine, so the network can't be flaky, so something's wrong, and after thinking about it, I realised this was the net/http.Server.IdleTimeout kicking in. Envoy maintains a connection pool to the upstream, but not every connection gets used before the IdleTimeout is up and reset, so occasionally, stale connections in the pool get killed, and envoy only realises when it tries to use them.

In this situation, with a trusted downstream (envoy) that has its own timeouts for dealing with untrusted clients, we can just not set the IdleTimeout and keep the connections around for longer.

Diagram updated

Anyway, here's a diagram of the timeouts, a few more control points have become available since Filippo's posts.

net/http timeouts