I don't think the article gives us the data to know this. Where did the latency spikes in the original implementation come from? Would fixing them have required a complete rewrite of the Python parts anyways?