Building a financial exchange like the one that LMAX Exchange runs is unique challenge. Customers demand low and predictable latency at ever increasing volumes. Since its launch 4 years ago it has grown to a system that regularly spikes beyond 10K tx/sec with latency measured in the 100s of microseconds.   Over that time at LMAX we've learnt a lot about software performance. Much to my disappointment it’s not all cache-line counters and bit twiddling. Many of our big wins have come from less obvious places. This talk will try to cover some lessons hard-learnt. Such as how to lie with numbers, why starting a solid foundation built from a clean domain model is imperative and understanding the necessary trade-offs between performance and safety/generality/functionality.   This talk will take a (slightly nostalgic) look through the evolution of the LMAX exchange and into the details of some our better decisions and attempt show how some of the ideas can be applied elsewhere.