Curiously, the express model ends up being a more clean style when creating long lived writer streaming endpoints, which is becoming much more popular with LLM’s.
All that was old will be new again.
All that was old will be new again.
I haven't done it though, I used to really like SSE (it's a little more REST-y) but the world seems to have decided on websockets.
Personally I like to strip all the excess data as soon as possible and expose the raw response as a plain “transfer-encoding: chunked” stream. That way the client can decide if they want to accept each chunk as they come or simply await the whole response with limited code change.
Either way, being able to easily pass the response writer around and set headers, write, close, as needed without being tied to a single function’s control flow is nice. It can be done in the more modern frameworks ofc, but ends up being a bit more machinery required at each endpoint.