Pure Nim Http2 client/server implementation. Tested with h2spec, and h2load.
Beware this library is in heavy development, and the API is not stable.
Note
This library supports HTTP/2 only. Not HTTP/1, nor HTTP/3.
nimble install hyperx
Nim +2.0
- OpenSSL
# this define needs to be in the main nim file
# or pass it as a compiler parameter `-d:ssl`
# or define it in the nim.cfg file
{.define: ssl.}
import std/asyncdispatch
import pkg/hyperx/client
proc main() {.async.} =
var client = newClient("www.google.com")
withClient(client):
let queries = ["john+wick", "winston", "ms+perkins"]
var tasks = newSeq[Future[Response]]()
for q in queries:
tasks.add client.get("/search?q=" & q)
let responses = await all(tasks)
for r in responses:
doAssert ":status: 200" in r.headers
doAssert "doctype" in r.text
waitFor main()
echo "ok"Beware HTTP/2 requires TLS, so if you want to test the server locally you'll need a local cert. I used mkcert to generate mine. Idk if there is an easier way to try this.
This will print received frames, and some other debugging messages
nim c -r -d:hyperxDebug client.nim
You may use a reverse proxy such a Caddy or a cloud offering such as AWS ALB for this. There are many ways to enumerate them all, but basically look for a service that can receive http/1, 2, 3 traffic and forward it as http/2.
Both server and client support data streaming. See the examples.
Backpressure is based on the http2 spec flow-control. The amount of data a stream can process at a time is bounded by the flow-control window size. The initial window size is 64KB. Note not calling recv in time will buffer up to window size of data.
Http/2 allows full-duplex data communication over a single stream. If you plan to only ever use this client and server, you won't need websockets.
Web-browsers only support half-duplex streaming for whatever reason, though.
If you need to support web-browsers it gets tricky for some use cases, as you need to do two fetches, one to stream the request, one to stream the response, and something like an ID in the URL so the server can associate both streams.
The CI runs h2load on it, but it only starts a single server instance. Proper bench-marking would start a server per CPU, and run a few combinations of load. To save you the trip, the CI results show around 30K requests/s.
MIT
