go-fastcounters

This demonstrates a very simple Go HTTP server that can:

• process 200k TPS at ~300microseconds latency
• with each message being persisted to disk before being processed
• all transactions processed serially and therefore atomically

The chess example shows 70k TPS at sub-millisecond latency. This is equivelant to about 5b chess games a month. Chess.com has around 1b games per month. So this is the kind of scale you can get if you keep the hot-path fast and move the slow stuff out to other services.

Run the benchmark

go run .
wrk http://127.0.0.1:8080/ --latency -t8 -c64 -d60s

The Code

The "framework" code (i.e. the code that writes to disk and orders all messages) is in wal/wal.go

The other files are:

• main.go - starts up the WAL and the http server
• counter_state.go - the processing of messages to update state
• counter_routes.go - the http endpoints

Similarly the state and routes files are available for the chess server as well.

The approach

By bringing all messages back to a single thread, they can be totally ordered and persisted to disk in batches. Then they are sent in this order to the message handlers that update state (also on a single thread).

Inspired by the LMAX Architecture