The Hot Path

You own the latency budget of a trading system for one day. Each interval is a component — the ingest buffer, the order map, the book's price levels, the top-k signal scan, the covariance step, the P&L accumulator, the parallel reduction — and for each one you choose a specification and set one number: a capacity, a tick range, a value of k, a thread count. Then a cost model charges you for the operations your choice performs and for the memory stalls they incur at whatever cache level your working set actually landed in. A hash map is right until the data leaves cache. An array by tick is right until the range is wide. Kahan summation is right only when the precision requirement asks for it. Every one of those depends on numbers you are shown, and they change.

⏱️

BudgetNanoseconds of headroom per message, banked across the day. Come in under budget and it builds; blow it and it burns. Zero → the system falls over.

📶

ThroughputThe share of offered messages you actually handled. Miss the budget and messages queue; miss it badly and they drop.

🎩

Desk standingHow close your specification came to the best one available on what you could see. Not the same as the latency. Zero → they take the service off you.

🧠

MemoryResident bytes against the box. Past 100% the process is killed and rebuilt, and the interval is lost.

Nothing is scripted. The workload is drawn, the cost model is real, and the best answer changes every interval.