I started to use elixir for hobby project (ad exchange) and stucked with performance of http clients (with ssl):
I get very low RPS (400) for requests with high latency (200-300msec) and high cpu usage (20% per core) on aws c5.large (2 cores) using elixir 1.6.4.
I tried different http clients (hackney, lhttpc, dlhttpc) with different pool sizes and still getting little improvement (lhttpc + dispcount gets the best numbers)
Is it OK for elixir/erlang http clients to have such low performance?
Maybe try https://github.com/lpgauth/buoy … That’s what I usually use when I need performance, but I don’t think I ever make requests with such a high latency.
Note that it doesn’t implement the full http/1.1 spec: https://github.com/lpgauth/buoy#unsupported-features.
But it is the most performant http/1.1 client afaik: https://github.com/lpgauth/httpc_bench.
I’ve often seen 200-300msec latency for HTTP client requests with httpc or hackney, so that’s about what I’d expect. I assume you’re using keep-alives, etc, as appropriate.
What kind of latency are you used to having from HTTP clients in other languages?
Edit: Or is it just the throughput that’s unexpected?
I’ve often seen 200-300msec latency for HTTP client requests
Oh, right. I wasn’t thinking straight. These are normal latencies indeed …
Thanks for suggesting! I tried buoy but encountered on strange behavior when remote endpoint became unresponsive –
shackle application which buoy using under the hood stops and all requests fail
which numbers do you get with buoy and on which server spec? I think that 400 RPS for just outgoing http requests without any logic looks odd for 2 cores
which numbers do you get with buoy and on which server spec?
I’d need to measure it. Sorry, don’t have any numbers right now. The server was a preemtible alpine vm on google cloud, and the requests from it were made to google storage api, so latencies were pretty low.