The issue

Typical server written in go has the following loop (net/http.Server has similar loop):

for {
    conn := listener.Accept()
    go serveConn(conn)
}

or

for {
    req := getRequest()
    go serveRequest(req)
}

This code starts new goroutine for serving each incoming connection or request. Go runtime starts new goroutines with small initial stack sizes (2KB). This stack may grow if serveConn or serveRequest calls many nested functions. This results in unnecessary CPU time spent in newStack calls for each incoming connection or request. See this article from Uber engineers for details.

The solution

Create new goroutines with dynamic initial stack sizes according to stack size stats from recently exited goroutines. This will eliminate newStack calls in typical server code, including net/http.Server.

The workaround I don't like

To serve connections / requests via goroutine pool. See this article for details.

cc'ing @josharian and @aclements