4.x: pooled/prewarmed TCP connections ignore connect_options (TCP_NODELAY lost), and hackney_tcp no longer sets nodelay #970

Version: hackney 4.8.5, Erlang/OTP 29.1.1, Elixir 1.20.4, Linux.

Summary

Requests that ask for socket options through connect_options (for example [nodelay: true]) can be served by pooled TCP connections that were opened without them:

1. Prewarmed connections ignore the caller's options. hackney_pool:prewarm_connections/5 opens connections with connect_options => [] and ssl_options => [] (hackney_pool.erl, around line 1375).
2. The plain TCP pool key does not include connect options. The key is connection_key(Host, Port, Transport) (hackney_pool.erl:131). So a prewarmed or earlier connection is checked out by a later request whose connect_options differ.

Why it matters: TCP_NODELAY

hackney_tcp:connect/4 uses [binary, {active, false}, {packet, raw}] as its base options, with no {nodelay, true}. hackney_ws, hackney_http_connect and hackney_socks5 do set {nodelay, true}.

After upgrading from hackney 1.25 to 4.x we saw:

• WebDriver calls (Wallaby to chromedriver on loopback) go from about 0.8 ms to about 83 ms.
• Every small POST with a body waits for Nagle plus delayed ACK: about 41 ms extra per request on loopback. 1.25 measured about 41 ms total on the same server.

We worked around it by passing connect_options: [nodelay: true] per client. That fixes fresh connections, but once a host has been prewarmed, the prewarmed connections are reused without nodelay and the delay comes back.

Reproduction (plain HTTP)

1. Start a local keep-alive HTTP/1.1 server.
2. Make a POST with connect_options: [nodelay: true]. It is fast.
3. Trigger prewarm for that host, or exceed the pool so the maintain-prewarm path runs.
4. Make the same POST again. It is served by a prewarmed connection and waits about 40 ms.

Suggested fixes, any of which would help

• Set {nodelay, true} by default in hackney_tcp:connect/4, as the websocket and proxy transports already do, and as 1.x behaved in our measurements.
• Carry the pool's or first request's connect_options into prewarm_connections.
• Include connect_options in the TCP pool key, the way the SSL path keys on TLS options, so connections opened with different socket options are not shared.

Happy to test a patch.