Sure, startup is defunct now and I think arbitrage & data on centralized exchanges is a dead market now. Wall Street HFTs got into the arbitrage game, and the data sites laypeople actually visit are the ones started in 2014.
Codebase was pure server-side Kotlin running on the JVM. Jackson for JSON parsing, when the exchange didn't provide their own client library (I used the native client libraries when they did). Think I used Undertow for exchange websockets, and Jetty for webserving & client websockets. Postgres for DB.
The threading model was actually the biggest bottleneck, and took a few tries to get right. I did JSON parsing and conversion to a common representation on the incoming IO thread. Then everything would get dumped into a big producer/consumer queue, and picked up by a per-CPU threadpool. Main thread handled price normalization (many crypto assets don't trade in USD, so you have to convert through BTC/ETH/USDT to get dollar prices), order book update, volume computations, opportunity detection, and other business logic. It also compared timestamps on incoming messages, and each new second, it'd aggregate the messages for that second (I only cared about historical data on a 1s basis) and hand them off to a separate DB thread. DB would do a big bulk insert every second; this is how I kept database writes below Postgres's QPS limit. Client websocket connections were handled internally within Jetty, which I think uses a threadpool and NIO.
Key architectural principles were 1) do everything in RAM - the RDS machine was the only one that touched disk, and writes to it were strictly throttled 2) throw away data as soon as you're done with it - I had a bunch of OOM issues by trying to put unparsed messages in the main producer/consumer queue rather than parsing and discarding them 3) aggregate & compute early - keep final requirements in mind and don't save raw data you don't need 4) separate blocking and non-blocking activities on different threads, preferring non-blocking whenever possible and 5) limit threads to only those activities that are actively doing work.
Would you use Kotlin again for the back end? Having not yet used it for that purpose, it seems like you’d get the benefit of the JVM ecosystem along with a nice language (but perhaps too many power-features).
For that particular use-case (or related financial ones) I'd consider Rust, which was a bit too immature when I was working on this but would give you some extra speed. HFT is winner-take-all and the bar has risen significantly even in the last couple years, so if I were putting my own money at risk now I'd want the absolute fastest processing possible.
i tried to do something similar recently. Your architecture sounds a lot like mine. I did all the thread pool management and JSON parsing using C++ libraries/frameworks.
The real-time data was visualized on a website. Here is an example.
https://algot.io/
