Show HN: Golongpoll – Go HTTP longpolling library
github.com> Why does everyone run to websockets even when they only need server-to-client pushing? Probably because it's difficult to get longpolling right.
Definitely not. Longpolling adds a huge amount of overhead when messages are sent relatively frequently. For each longpoll request, the client needs to send all the headers and cookies every time. Each response includes all headers and connection setup. A websocket connection has zero overhead after the initial http upgrade. If you have websocket available, you should always prefer websocket to long polling. In most cases where websocket would be blocked by infrastructure, long polling would be forced to retry extremely quickly because of forced timeout.
WebSockets are superior from a technical perspective, no arguing that. But, long-polling makes for more straightforward APIs, which could be valuable if you're more concerned about developer UX than the number of bytes on the wire.
For example, long-polling can actually be RESTful, so it fits in well with conventional API practices. It's also naturally resilient to network issues since each re-poll heals the stream. Users are very unlikely to lose data due to forgetting about some corner case.
Websockets can be RESTful too. I've seen the following:
The problem with this type of RESTful integration though is the connection limit imposed by browsers. Of course, long poll has the identical problem. You can't long poll more than a couple endpoints.GET /orders - fetch all orders GET /orders Upgrade: websocket - streaming view of all new ordersTrue, you can use URI paths with WebSockets which at least makes the streams resource-oriented. CBIX does it: https://www.cbix.ca/api-websocket
I think calling this RESTful is a bit of a stretch though, at least in any conventional sense of what that means to people.
There's also EventSource, which I've had some luck with. It's more REST-y, I find.
Long polling is a hack. Anyone who has had to scale it across more than a single CPU core on a single machine will know this.
TCP and WebSockets are stateful, HTTP is stateless. So long polling is essentially:
Stateful -> Stateless -> Sateful
You start in a good place, then you add complexity to make it stateless and then you add even more complexity to roll back the statelessness and then you end up where you started... Except it's 90% slower and you now have to use an armada of sticky load balancers to prevent the system from falling apart...
Maybe in "fat" language runtimes this is a problem (Ruby, Java, Python etc).
In Go, requests are a single goroutine, dirt cheap (no problem with a single CPU). To me, this type of work is where Go excels.
I used SSE (not quite long polling but I think it has the same implications as far as this discussion goes) back in Go 1.3 days (so GOMAXPROCS=1) to write metrics to 100<n<1000 clients and the process barely registered CPU or memory usage. No lock problems either (copying metric values over channels).
As far as complexity .. well, sure, if you fuck up the design or use the wrong tool for the job you'll have problems.
When I needed some simple push notifications, I could't find any decent library or example for longpolling in go (or most other languages), so I decided to make one :)
A lot of the examples out there in blog posts are too simplistic. They don't handle buffering old events and most of them simply give an event to the first request and then discard it. That's not real pub-sub.
I'm pretty sure my blog post with a long polling tutorial was one of those. Glad to see someone improve on the idea and turn it into a lib!
Great to see this. I am a big fan of long-polling. Too many people only think of it as some hack underlying libraries like Socket.io or SockJS, but this is because they've not seen how clean a long-polling API can actually be.
Yeah. I originally started out using SSE but I ended up settling on longpolling because of wider browser support and the fact that there are issues with SSE even on newer browsers and polyfills. The knee-jerk reaction nowadays is to use websockets for everything, but sometimes simple is better.
I've been using SSE with Go for personal projects, so just one browser, could you expand on the issues you mentioned? I'm curious if it is indeed as bad as you say.
Doesn't work in IE, last I checked.