While implementing Apache Arrow for Erlang recently, I found handling the intensive parts in native code slow, so I wrote it in pure Erlang. It had to be performant because it would support Erlang's Telemetry implementation - vital to any BEAM web app.

With my initial code too slow, I had to learn a black art: optimizing sequential Erlang.

But it was easy to find optimizations to make, even easier to over optimize, and painfully easy to optimize the wrong thing completely. I realized that I had 2 aspects to learn:

First, making a single optimization: How do I choose the right knob to turn? By how much? Did I get the results I expected? Is it maintainable?

Second, optimizing the system as a whole: Which bottlenecks to focus on and by how much? How do I stay pragmatic?

Let me share what I learned:

- The processes of performance optimization.
- Applying CS theory to design optimizations.
- The specifics of profiling, optimizing and hypothesizing for the BEAM