The hot path of a triangular-arbitrage bot without the orders: 1,375 pairs, 3,230 cycles, a seqlock per symbol, integer-only evaluation, one thread per connection and no allocation after startup. What it costs, measured, and the tests that keep it honest.
Oct 13, 2026 · 13 min readMeasured from AWS Tokyo with a chrony-synced clock: the gap between the matching engine and the public stamp, the 0.9 ms publisher floor, what extra connections and backends can and cannot buy, FIX against WebSocket, and why the public feed lands 0.6–0.8 ms after the people who act on it.
Oct 13, 2026 · 14 min readA paper-trading simulator over a per-symbol quote history, run from home and from Tokyo: zero fills at any latency. Then the recording that explained it: the signals live about a millisecond, somebody takes 81 % of them 0.8 ms after the public stamp, and the arithmetic of where their data must come from.
Oct 13, 2026 · 15 min readIf the taker loses because a sloppy resting order gets picked off, be the resting order on purpose, priced from the liquid pair and hedged fast. Replayed over two hours of trades and quotes on 36 thin pairs: negative at every edge, negative at zero latency, and negative even at the front of the queue. Where the money goes, per fill.
Oct 13, 2026 · 10 min read