Broken pipe on unix socket error handle worker: error send AgentDisconnect frame

[alechenninger]

When listening on a unix socket, I get this error a few seconds after each event is processed:

2020/12/14 18:37:58 error handle worker: error send AgentDisconnect frame: write unix /var/run/gohaproxy/agent.sock->@: write: broken pipe

Example code https://github.com/alechenninger/go-haproxy-agent/blob/cd1d9681bcfe818f134dc1bc2d2ba1ec4e5983f2/main.go#L43

I've tried both listneing on "unix" and "unix packet"

Haproxy configs here https://github.com/alechenninger/go-haproxy-agent/tree/cd1d9681bcfe818f134dc1bc2d2ba1ec4e5983f2/conf

[alechenninger]

Some additional context: I'm seeing this on mac while using docker for mac, and a shared unix socket between containers. Docker for Mac doesn't support unix sockets from the mac OS host, but it does in between containers.

I tried switching to https://github.com/criteo/haproxy-spoe-go and am seeing the same results, so wondering if this is a product of my environment.

[alechenninger]

Did a bit more experimenting. Looks like timeout handling is different with unix sockets vs tcp.

With TCP, server and idle timeouts don't trigger errors.

With unix sockets, server and idle timeouts do trigger errors. This affects both this and the criteo libraries.

It seems benign, and may just warrant higher timeout values to reduce noise.

[negasus]

Thank you for such a detailed description. Can I close this issue?

[alechenninger]

It seems like a bug that one protocol emits error logs when another doesn't. It may be misleading, because AFAICT these aren't really meaningful "errors" in the sense that nothing appears to be actually wrong. Ultimately, whether you leave the issue open is up to you.