One full NASDAQ trading day of AAPL, replayed through the book

Methodology & caveats

1. ¹ Latency is the wall time of Book::apply(event) measured per event with std::chrono::steady_clock, whose tick on Apple silicon is 41.67 ns (24 MHz) — values are quantized to that grid and include roughly one tick of timer overhead. Percentiles are over all 4,772,682 events.
2. The playback is a faithful re-broadcast of the engine's recorded output: every top-of-book price change (millisecond-timestamped) and every fill are replayed exactly; only the ladder's interior levels refresh on a 2-second heartbeat. Speeds are market-time multipliers — the engine itself consumed the day in 0.42 s.
3. This is the v0 price-level book: an order-ID hash map plus a sorted map of occupied price levels per side. Per-order FIFO queues and matching come next, so the metric reported is book-update latency, not order-to-fill.
4. Data: Databento XNAS.ITCH MBO (market-by-order) for AAPL on 2024-01-05, DBN-encoded. Parser record counts match the Python databento client exactly (2,258,537 adds · 2,279,209 cancels · 106,477 fills · 128,458 non-book trades). Trade direction in time & sales is inferred from the resting side of each fill.
5. Regular trading hours are 09:30–16:00 ET; the replay loops over that window. The 16:00 spike in the message-rate chart is the closing cross.