cowboy_http holds onto memory with long-lived connections

[rlipscombe]

We noticed a problem last week when using long-lived client connections: The process running cowboy_http:loop/1 doesn't GC after sending the response body to the client.

In some cases, this can result in holding on to (e.g.) 10MiB or so of garbage before either the client sends another request, or reaching idle_timeout.

Would it be possible to trigger a GC after the response body has been sent? This would be similar to gen_server's hibernate, or the ssl application's hibernate_after option.

Obviously, we could set idle_timeout more aggressively, but that closes the connection, which seems to defeat the purpose of long-lived connections.

I can reproduce this against one of our internal cowboy-using applications by using some relatively simple gun client code. My example code uses HTTP/1.x, but we've seen this with (I think) a Go client using HTTP/2.

(cowboy 2.9.0, ranch 1.8.0)

[essen]

We can introduce a way to hibernate by sending it a message (via cowboy_req:cast) or similar.

[essen]

I guess it could hibernate automatically via an option as well after all messages if you don't expect much to happen. Option could be set both at startup and dynamically.