Why Trading Needs Its Own Special Network
Low-Latency HFT Network Engineer · Module 1: Latency From Zero
Lesson 5 of 7
Prerequisites: The Internet's Plumbing: Where Packets Actually Go, Why Microseconds Are Money, What Trading Is (and Why Being Fastest Wins)
What you'll be able to do: look at two latency measurements and pick the link a trading firm would actually buy — and explain why the lower average can be the worse choice.
Think about your morning commute. Some days you hit every green light and arrive in ten minutes; other days a garbage truck, a red light, a detour — twenty-five. Same road, wildly different trip. Now imagine your boss pays you only if you arrive within eleven minutes, every single day, and a minute late costs you your job. You wouldn't take the city street. You'd build a private road with no lights, no trucks, no surprises — and you'd drive a stopwatch on it daily. A trading firm lives that commute. It sends buy and sell orders racing to the exchange, and arriving a fraction of a second late means losing the trade — thousands of times a day. Regular internet is the city street: shared, unpredictable, fine for email. Trading needs its own private road: built for one job, measured constantly, never a surprise. Here's the puzzle.
Scenario. A trading firm is choosing which of two network links will carry its real trades to the exchange. An engineer sends a test message across each link 10 times and times every trip. Link B's average looks better — but the engineer refuses to sign off on it. The firm will route millions of dollars of trades through whichever link wins, so the decision has to be right, not just fast-looking.
Given artifacts. The test rig and the two measurement exhibits:
Exhibit 1 — Link A, 10 trips (microseconds): 48, 49, 50, 50, 51, 52, 48, 51, 50, 51
Exhibit 2 — Link B, 10 trips (microseconds): 5, 8, 6, 200, 7, 5, 150, 6, 5, 9
Your task: name which link the trading firm picks; compute the average trip time and the worst-case (slowest) trip time for each link, showing your math; and write one sentence explaining why the average lies.
Workspace: analyze-and-answer — three short text boxes: (1) your pick (Link A or Link B), (2) your averages and worst cases with the arithmetic shown, (3) your one sentence on why the average lies. Nothing is graded; the boxes simply record your attempt before the worked answer.
Hint 1 — where to look
The answer is hiding in the two columns of numbers — but don't stop at the average. Look at the single slowest trip in each column and ask what that trip would cost in a real trading day.Hint 2 — what to compare
Compute each link's average and its worst (slowest) trip. Then ask the question a trading firm actually asks: which link's worst trip could you afford to live with, thousands of times a day?Hint 3 — the mechanism
A trading firm's rule is "plan for the worst trip, not the average one" — because the trip that arrives late is the trip that loses the trade. Run both sets of numbers and apply that rule; the exhibits contain everything you need.Commitment ritual: below the workspace sits a checkbox — "I've attempted this challenge and thought it through." Checking it (with or without typing an answer) reveals the worked answer in S7. Honor system: the page hides the answer until you commit. The struggle is where the learning happens.
Checking the box reveals the worked answer in S7 below. Returning learners stay unlocked.
Latency is the trip; jitter is the wobble
Latency (how long a single message takes to travel from sender to receiver) is what Lessons 1–4 measured. But one trip tells you almost nothing. Send ten messages and the trip times form a spread — and that spread has its own name: jitter (how much the travel time changes from one message to the next).
Picture a metronome versus a drumroll. A metronome clicks every second, exactly — low jitter. A drumroll is a burst of hits with uneven gaps — high jitter, even if the average gap is the same. Link A in the challenge is the metronome; Link B is the drumroll.
Why this matters for the challenge: the two exhibits differ far more in their wobble than in their averages — and we'll solve which one a firm can actually plan around in the worked answer.
Why shared networks wobble
Your office network is a shared road. Dozens of people stream, download, and video-call over the same links, and their traffic arrives in bursts (sudden clumps of messages, then silence). Inside every network box sits a queue (a waiting line where messages sit when the box is busy). When a burst hits, your message waits in line — and waiting time depends on how big the burst was, which is different every time.
That waiting is congestion (too much traffic arriving at once, forcing messages to queue). Congestion is the main factory that manufactures jitter: the same message, the same road, a different wait every trip. An office network is designed for sharing fairly, not for arriving on time.
Why this matters for the challenge: those 200 and 150 microsecond trips on Link B are what congestion looks like from the outside — rare, violent, and invisible in the average.
Determinism: the real prize
Here's the word trading firms actually optimize for: determinism (every trip taking nearly the same time — predictability you can plan around). Given the choice between "every trip takes 50 microseconds" and "trips take 5, 200, or 8 microseconds at random," a trading firm takes the steady 50 every time.
Why? Because a trading strategy is a plan, and a plan needs a reliable clock. If you know every trip takes 50 microseconds, you can decide exactly when to send. If trips take anywhere from 5 to 200, you must assume the worst every time — and the "fast" link is secretly the slow one, because you're always budgeting for its worst mood. Determinism turns the network from a gamble into a measuring instrument.
Why this matters for the challenge: the firm's decision rule is "pick the link whose worst trip I can live with" — the worked answer applies it to both exhibits.
The third requirement: nothing may vanish
There's one failure worse than a slow trip: a trip that never arrives. Packet loss (messages that vanish inside the network and never reach the other end) forces the sender to notice the silence and send the message again — a whole second trip, with all its delay and all its jitter.
Trading networks are engineered to be effectively loss-free: enough capacity everywhere that queues never overflow, so no message is ever dropped for lack of room. Loss-free isn't a luxury; it's what keeps the first two goals (low latency, low jitter) from collapsing the moment traffic spikes.
Why this matters for the challenge: a single lost message would look like an infinitely slow trip — the ultimate worst case, and the reason firms measure loss as carefully as speed.
Why "fast internet" isn't enough
Your home fiber might advertise enormous bandwidth (how much data a link can carry per second — the width of the road, not the speed of the trip). Bandwidth is not the problem. Regular internet fails trading on the other three counts: it's shared (so it jitters), it's best-effort (the network tries to deliver, but promises nothing about when), nobody measures every trip, and your message crosses many shared boxes you don't control.
A trading network is the opposite in all four ways: private (nothing shares the road), engineered for determinism (same trip time, every time), loss-free (room for every message), and measured constantly (every trip timed, because what isn't measured can't be trusted). That's the private race track from the hook — not a faster internet, a different kind of network.
Why this matters for the challenge: Link A and Link B aren't "fast" and "slow" — they're two different kinds of behavior, and only one kind belongs on a race track. The worked answer names the winner.
- Step 1 of 5: Both networks send the same stream of messages — one after another, evenly spaced at the sender.
- Step 2 of 5: On the shared office network, each message collides with bursts of everyone else's traffic, so every trip takes a different amount of time — some sprint, some stall in queues.
- Step 3 of 5: Watch the orange arrival markers: they land in clumps and gaps instead of evenly. That wobble in arrival times is jitter, made visible.
- Step 4 of 5: On the private trading network, nothing shares the road, so every message takes nearly the same time — the green dots cross at a constant pace.
- Step 5 of 5: The green arrival markers land like a metronome: steady, predictable. That predictability — determinism — is the real prize, worth more than raw speed.
🔒 Revealed after the commitment ritual in S2 — attempt the challenge first. (Honor system: the page hides this until you check the box.)
Step 1 — do the arithmetic. Add each column and divide by 10:
- Link A: 48 + 49 + 50 + 50 + 51 + 52 + 48 + 51 + 50 + 51 = 500 → average 50.0 μs; worst (slowest) trip = 52 μs.
- Link B: 5 + 8 + 6 + 200 + 7 + 5 + 150 + 6 + 5 + 9 = 401 → average 40.1 μs; worst trip = 200 μs.
Step 2 — apply the firm's rule. The firm plans for the worst trip, not the average. Link A's worst trip is 52 μs — barely above its average. Link B's average is 10 μs better, but its worst trip is four times its average, and there are two of them (200 and 150 μs). The firm picks Link A.
Step 3 — why the average lies. The average lets nine fast trips outvote two catastrophic ones — but in trading, the two slow trips are the ones that lose the trades, and the nine fast ones don't compensate. An average describes the typical trip; the worst case describes the trip that costs you money.
Wrong turns, named. You might have picked Link B because 40.1 < 50.0 — that's the average seducing you; the exhibit's two spikes are the evidence against it. You might have picked B for its 5 μs best trip — but a firm can't plan around a best case any more than an average; only the worst case is plannable. You might have dismissed the spikes as "rare" — but a firm sends thousands of messages a day, so a 2-in-10 event becomes hundreds of lost trades a day.
Verify it worked: re-run the test with 1,000 trips instead of 10 and compare the worst trip of each link. Link A's worst will sit near ~52 μs; Link B's worst will grow — the spikes get worse with more samples, never better. If a longer test ever shows B's worst trip beating A's, the pick was wrong.
Check yourself — nothing here is graded. Wrong answers are the useful ones; each explains why.
Question 1. One message takes 48 μs, the next takes 52 μs, the next takes 49 μs. Which statement is correct?
Question 2. Your firm tests two connections for a week. Cable X: every trip lands between 40 and 44 μs. Cable Y: trips land between 8 and 180 μs, averaging 30 μs. Which cable carries live trades, and why?
Question 3. A friend says: 'My home fiber is super fast — 1,000 megabits per second! A trading firm should just use that.' What is the flaw?
Question 4. Put these in the order they happen when a burst of other traffic hits a shared network:
- Jitter is the variation in trip time from one message to the next, and a trading firm fears jitter more than a slightly slower steady link.
- An average lets many fast trips outvote a few catastrophic ones — always demand the worst case next to the average.
- Determinism means every trip takes nearly the same time, which turns the network from a gamble into a plannable instrument.
- Shared networks wobble because bursts of other people's traffic force your messages to wait in queues.
- A trading network is private, deterministic, loss-free, and constantly measured — "fast internet" is none of those four things.
Next: The Cost of a Millisecond: Real Disasters — you've just learned what a trading network must be; now see what happens when a firm has the speed but forgets the safety, in two real disasters measured in millions of dollars per minute.