My current leading theory is this:
Starting node.js 20, requests try to connect to both IPv4 and IPv6 addresses, and use the socket that opens first.
So, this bug exists only for services that resolve their domain to IPv6, but somehow aren’t actually responding to those v6 addresses in a timely manner.
I’ve updated our base docker image to the latest version of node.js 20, so that if this is somehow fixed in there already, the next release should hopefully also fix this.
However if the issue still persists in the next release, I’ll create a quick PR to disable the new socket autoSelectFamily thing, and create a new patch release right away.