Design PFX: Fairness engineering

Lesson 4.6

Fairness engineering

What you'll learn

The mechanisms that make equal speed an engineered guarantee rather than a hope — equalized paths, fair gateways, honest ordering — and the market designs that make speed matter less at all.

Building on

4.1 framed latency as a tax. This lesson is about who pays it — because an exchange that is fast on average but faster for some is worse than one that is uniformly slower. Fairness, not speed, was always the regulatory and moral requirement (0.1).

Inside the wire-to-wire window, unfairness hides in three places, each with an engineering answer:

flowchart TB
    PA["Participant A — rack 5 m from the exchange switch"] -- "fiber coiled to exactly 250 m" --> SW["Exchange access switch"]
    PB["Participant B — rack 200 m away"] -- "fiber run of exactly 250 m" --> SW
    SW --> TS["Hardware timestamp at first touch — THE finish line: arrival order fixed here, auditable forever (4.2)"]
    TS --> GWQ["Gateways poll all sessions round-robin — one firm's message flood cannot bury another's single cancel (2.7)"]
    GWQ --> SEQ["Sequencer processes in stamped arrival order — no reordering downstream (1.7)"]
    MDOUT["Market data egress — released to all ports simultaneously, cable lengths equalized again (0.1)"]
    SEQ -.-> MDOUT

Fairness as physical engineering. Colocation halls literally coil fiber so every participant's cable is the same length — a nanosecond-level guarantee no policy document can match. The finish line (hardware timestamp at first touch) is placed before anything shared, so no queue inside the exchange can reorder the race.

Equalized ingress. Same cable length, same switch generation, same port speed for every colo participant. Everything before the timestamp is the race the exchange promised to referee impartially; everything after it is FIFO by stamp. Note why timestamping alone isn't enough: the stamp must be taken at a point participants reach fairly — a timestamp after an unfair queue merely notarizes the unfairness (0.2's audit-trail obligation, applied to nanoseconds).
Fair sharing inside. Round-robin polling across sessions at gateways, per-participant throttles (1.6, 2.7) — one member's message storm degrades their own queue, not the hall's. The 2.7 overload ladder was fairness engineering all along; this lesson just lowers its resolution from milliseconds to microseconds.
Simultaneous egress. Market data (0.1's "at once" requirement) leaves all ports in the same transmission cycle, cables equalized again — and private fills must not systematically precede the public print, or the counterparty's counterparty learns the price late (1.2's private/public twin rule, now with a stopwatch).

Then there is the deeper move: changing the game so speed buys less. Two designs worth knowing, both real. A speed bump — a fixed delay (IEX's famous 350 µs coil of fiber) applied to all incoming orders, giving the venue's own systems time to update reference prices before fast traders can exploit stale ones: latency arbitrage's profit window is surgically removed while everyone's relative speed is preserved. And frequent batch auctions — run the 1.6 auction mechanism every 100 ms instead of matching continuously: within each batch, arrival order doesn't matter at all, so the nanosecond arms race becomes economically worthless by construction. Both are the 3.1 mindset applied to speed: ask what the contract actually requires — fair price formation — and spend engineering there, not on the arms race's scoreboard.

Key ideas

Equalize ingress, share fairly inside, release simultaneously — and place the finish line before anything shared.
Timestamps notarize; only physical equalization makes the race fair to begin with.
Speed bumps and batch auctions redesign the game so nanoseconds stop paying — fairness by market design.

Fairness engineering

See what actually stuck.