---
title: "Buying Connectivity"
book: "Networks, Hardware and Trading Infrastructure"
subject: quant
language: en
chapter: 15
exercises: 8
source: https://one-course.com/books/quant/14/en/chapter/15-buying-connectivity
---

# Chapter 15 — Buying Connectivity

A cloud provider’s [service-level agreement](#def-nw-buying-connectivity-sla) for its private connections promises 99.99% of each month to customers who buy two connections at two sites, and a credit of 10% of the month’s port charges if it fails to deliver. For one connection it promises 95%, and pays nothing unless the month has lost more than a day and a half. Trading firms buy circuits on the same kind of terms: the provider promises a percentage, pays a fraction of a monthly fee when it misses, and the loss from the outage (the positions that could not be hedged, the quotes that could not be pulled) is the firm’s. This chapter is about what a firm buys, from whom, on which terms, and why the second circuit is worth buying only if it does not share a trench with the first.

Chapters 9 to 14 described the buildings, the maps and the routes. Here they become purchases: [cross-connects](https://one-course.com/books/quant/14/en/chapter/9-colocation-products-and-how-they-are-sold#def-nw-colocation-products-and-how-they-are-sold-mmr) and circuits in a building, [metro circuits](#def-nw-buying-connectivity-metro) between buildings, long-haul fibre and radio between cities, and feeds delivered by others. The prices that are public come from venues’ fee schedules; the rest are negotiated, and the chapter says so.

## 15.1 What is bought: circuits, wavelengths, bandwidth and feeds

**Definition 15.1 (Metro circuit).**

A *metro circuit* is a point-to-point connection between two buildings in the same metropolitan area (two data centres, a data centre and an office), sold as a [lit service](https://one-course.com/books/quant/14/en/chapter/13-long-haul-fibre#def-nw-long-haul-fibre-dark) of a given speed or built on leased or owned fibre.

A firm’s connectivity is a stack of purchases. In each building it buys [cross-connects](https://one-course.com/books/quant/14/en/chapter/9-colocation-products-and-how-they-are-sold#def-nw-colocation-products-and-how-they-are-sold-mmr) (chapter 9) and the venue’s own ports; between the buildings of a metropolitan area, [metro circuits](#def-nw-buying-connectivity-metro); between cities, wavelengths or [dark fibre](https://one-course.com/books/quant/14/en/chapter/13-long-haul-fibre#def-nw-long-haul-fibre-dark) (chapter 13) and radio bandwidth (chapter 14); and, for data it does not collect itself, feeds delivered by an extranet or a vendor. A New Jersey firm that trades the NYSE, Nasdaq and Cboe from one building buys at least two [metro circuits](#def-nw-buying-connectivity-metro) and the venues’ ports at the other end; one that trades Chicago futures against them adds a long-haul route and its back-up.

![A New Jersey firm’s connectivity as purchases, schematically: cross-connects and ports in its own building, metro circuits to the other two buildings of the triangle (chapter 10), a long-haul route with its back-up to Chicago (chapters 13 and 14), and a circuit to a provider for the data it does not collect itself.](https://one-course.com/images/onecourse/chapters/quant-14/nw-buying-connectivity/fig-b344651bec77.svg)

***Figure 15.1.** A New Jersey firm’s connectivity as purchases, schematically: [cross-connects](https://one-course.com/books/quant/14/en/chapter/9-colocation-products-and-how-they-are-sold#def-nw-colocation-products-and-how-they-are-sold-mmr) and ports in its own building, [metro circuits](#def-nw-buying-connectivity-metro) to the other two buildings of the triangle (chapter 10), a long-haul route with its back-up to Chicago (chapters 13 and 14), and a circuit to a provider for the data it does not collect itself.*

**As of September 2026 — Published connectivity prices: orders of magnitude.**

The NYSE group’s connectivity fee schedule (last updated April 2025) charges, per month: 2 500 USD for a 1-gigabit IP-network circuit, 11 000 and 18 000 USD for 10- and 40-gigabit IP and NMS network connections, 600 USD for a fibre [cross-connect](https://one-course.com/books/quant/14/en/chapter/9-colocation-products-and-how-they-are-sold#def-nw-colocation-products-and-how-they-are-sold-mmr) in the data centre, 9 000 to 44 000 USD for 10 to 200 megabits of wireless bandwidth between Mahwah and Secaucus (10 000 to 45 000 to Carteret), and 6 000 USD for a wireless connection carrying CME Group data; one-time charges of 500 to 15 000 USD apply. MIAX Pearl charges 15 000 USD a month for a 10-gigabit ultra-low-latency connection (chapter 9); Deutsche Börse 6 000 to 8 400 EUR a month for its 10-gigabit co-location connections (chapter 11).

| Item | Monthly | One-time | Source |
| --- | --- | --- | --- |
|  | (USD) | (USD) |  |
| NYSE IP network, 1 Gb circuit | 2 500 | 2 500 | fee schedule |
| NYSE IP and NMS networks, 10 Gb | 11 000 | 10 000 | fee schedule |
| NYSE IP and NMS networks, 40 Gb | 18 000 | 10 000 | fee schedule |
| NYSE data-centre fibre cross connect | 600 | 500 | fee schedule |
| NYSE wireless Mahwah–Secaucus, 10 Mb | 9 000 | 10 000 | fee schedule |
| NYSE wireless Mahwah–Secaucus, 200 Mb | 44 000 | 10 000 | fee schedule |
| NYSE wireless connection of CME data | 6 000 | 5 000 | fee schedule |
| MIAX Pearl 10 Gb ultra-low-latency | 15 000 | 0 | fee filing |

***Table 15.1.** The catalogue of `firm.slamodel`: published prices only, each with its ledger row and date. Wireless bandwidth costs about as much per megabit a month as a 10-gigabit fibre connection costs per gigabit. Data: `nw_buy.price_rows()`.*

The table is the scale of the market. A megabit of radio between two New Jersey buildings costs about 900 USD a month at 10 megabits and 220 USD at 200; a gigabit of fibre between a cabinet and the venue’s network costs about 1 100 USD a month at 10 gigabits. Speed, not bandwidth, is what the radio sells.

## 15.2 Financial extranets and managed infrastructure

**Definition 15.2 (Financial extranet).**

A *financial extranet* is a private network, run by a provider connected to many venues and firms, over which its customers reach venues’ market data and order entry, and each other, without building their own circuits to each; venues admit such providers under an agreement and charge them for the access they resell.

Nasdaq’s definition shows the arrangement from the venue’s side: an extranet provider signs Nasdaq’s extranet provider agreement, has its own connection to Nasdaq, is co-located in a Nasdaq facility, and gives its own customers access to Nasdaq’s services, in the building or elsewhere. For a firm the extranet is a trade: one connection to the provider replaces many to the venues, at the price of the provider’s latency and of sharing its network with its other customers.

**Definition 15.3 (Managed infrastructure).**

*Managed infrastructure* is equipment and connectivity that a provider installs, owns or leases, and operates on a firm’s behalf (servers, switches, circuits, market-data handlers), usually in the provider’s racks near the venues, sold as a monthly service.

[Managed infrastructure](#def-nw-buying-connectivity-managed) is the build-or-buy question of One Quant Book 16 (build against buy) at the scale of a rack: the firm that buys it gets a presence at a venue in weeks instead of months, and hands the provider control over the latency of every piece in the path. Firms whose edge is latency buy the building’s space and build the rest; firms whose edge is elsewhere buy the rest too.

## 15.3 Wireless-bandwidth vendors

Radio routes are sold in three ways. Venues sell bandwidth or data on their own links: ICE between its Mahwah data centre and the other New Jersey buildings (with Anova Financial Networks) and from Toronto; SIX between Zurich and Frankfurt; Euronext from London to Bergamo with McKay Brothers. Network operators sell bandwidth on theirs. And data vendors sell a market-data feed delivered over radio, as Quincy Data did from Aurora in chapter 10. What they have in common is scarcity: megabits, not gigabits; a fixed set of end points; and a fibre back-up that the firm must also buy, or accept to do without when it rains.

## 15.4 Contracts and service levels

**Definition 15.4 (Service-level agreement, service credit).**

A *service-level agreement* (SLA) is the part of a supply contract that states the performance the provider commits to (availability, latency, repair time), how it is measured, and what the provider owes when it fails. A *service credit* is what it owes: usually a percentage of the month’s charges for the failed service, credited against future bills.

**Definition 15.5 (Mean time between failures, mean time to repair).**

The *mean time between failures* (MTBF) of a component is the average time it runs between failures; its *mean time to repair* (MTTR) is the average time from a failure to its restoration. Its steady-state availability is $\mathrm{MTBF}/(\mathrm{MTBF}+\mathrm{MTTR})$.

**As of September 2026 — A published credit schedule.**

AWS’s [service-level agreement](#def-nw-buying-connectivity-sla) for Direct Connect (last updated July 2026) makes three commitments: 99.99% monthly uptime for multi-site redundant deployments, 99.9% for multi-site non-redundant ones, and 95.0% for a single connection. Missing them earns a credit of 10%, 25% or 100% of the affected connections’ port charges, at thresholds of (99.99, 99.0, 95.0), (99.9, 99.0, 95.0) and (95.0, 92.5, 90.0)% respectively. Credits are the sole remedy, and single connections are not recommended for production.

**Definition 15.6 (Route diversity).**

*Route diversity* is the property of two or more circuits between the same end points that they share no physical element whose failure would stop both: no duct, cable, building entry, room, piece of equipment or power supply.

**Proposition 15.7 (Two circuits).**

Let each of two circuits be unavailable a fraction $u$ of the time, independently, and let common-mode events that stop both occur a fraction $c$ of the time. The pair is unavailable a fraction $u^2 + c - u^2 c \approx u^2 + c$. When $c$ is of the order of $u$, the second circuit gains almost nothing; only diversity, $c \to 0$, delivers $u^2$.

**Proof.** The pair is down when both are down independently (probability $u^2$) or a common-mode event is under way ($c$); inclusion–exclusion gives the union. ∎

```python
def design_availability(design):
    """Down when every circuit is down (independently) or a common-mode event is under way."""
    p_all = 1.0
    for c in design.circuits:
        p_all *= 1 - availability(c.mtbf_h, c.mttr_h)
    p_common = design.common_rate_y * design.common_mttr_h / HOURS_Y
    return 1 - (p_all + p_common - p_all * p_common)
```

***Listing 15.1.** A design’s steady-state availability: every circuit down at once, or a common-mode event. code/firm/slamodel/firm_slamodel.py*

![Three designs, schematically: one circuit; two circuits whose fibres share a duct, so that one event (a cut, a fire) stops both; and two diverse circuits that share nothing between the firm’s and the venue’s rooms.](https://one-course.com/images/onecourse/chapters/quant-14/nw-buying-connectivity/fig-197b07be0675.svg)

***Figure 15.2.** Three designs, schematically: one circuit; two circuits whose fibres share a duct, so that one event (a cut, a fire) stops both; and two diverse circuits that share nothing between the firm’s and the venue’s rooms.*

The model gives each circuit two failures a year with a mean repair of four hours (MTBF $4383\,\mathrm{h}$, MTTR $4\,\mathrm{h}$), and the shared duct half a common-mode event a year with a mean repair of twelve hours (a cut fibre is spliced, not swapped): assumptions, labelled as such. One circuit is then down 0.091% of the year, 480 minutes. Two in one duct are down 360 minutes: the duct’s six hours a year, almost all of it, since two independent failures coincide for less than half a minute a year. Two diverse circuits are down 0.4 minutes a year.

| Design | Availability | Expected | 95th pct. | Annual cost | Loss at 2 000 |
| --- | --- | --- | --- | --- | --- |
|  | (%) | (min/yr) | (min/yr) | (kUSD) | USD/min (kUSD) |
| One circuit | 99.9088 | 479.6 | 1 541 | 139.2 | 959 |
| Two circuits in one duct | 99.9315 | 360.4 | 1 992 | 278.4 | 721 |
| Two diverse circuits | 99.99992 | 0.4 | 0 | 320.2 | 0.9 |

***Table 15.2.** Simulation of three designs of 10-gigabit NYSE connections (published prices; the failure rates, the duct’s common-mode events, a 15% premium for diversity and the trading loss per minute are assumptions): steady-state availability and expected minutes down, the 95th percentile of a simulated year (1 000 years), annual cost with one [cross-connect](https://one-course.com/books/quant/14/en/chapter/9-colocation-products-and-how-they-are-sold#def-nw-colocation-products-and-how-they-are-sold-mmr) per circuit, and the expected trading loss. Data: `nw_buy.results()`.*

![Expected minutes down a year against annual cost for the three designs of : the second circuit in the same duct doubles the cost and removes a quarter of the downtime; diversity adds 15% and removes nearly all of it. Data: fig_buy.py.](https://one-course.com/images/onecourse/chapters/quant-14/nw-buying-connectivity/fig-3d00b82f8462.svg)

***Figure 15.3.** Expected minutes down a year against annual cost for the three designs of [Table 15.2](#tab-nw-buying-connectivity-designs): the second circuit in the same duct doubles the cost and removes a quarter of the downtime; diversity adds 15% and removes nearly all of it. Data: `fig_buy.py`.*

```python
def simulate_year(design, seed=0, sigma=1.0):
    rng = np.random.default_rng(seed)
    per = [_merge(_outages(rng, 1 / c.mtbf_h, c.mttr_h, sigma)) for c in design.circuits]
    common = _merge(_outages(rng, design.common_rate_y / HOURS_Y, design.common_mttr_h, sigma))
    all_down = per[0]
    for p in per[1:]:
        all_down = _intersect(all_down, p)
    down = _merge(all_down + common)
    per_min = [60 * sum(b - a for a, b in _merge(p + common)) for p in per]
    return {"down_min": 60 * sum(b - a for a, b in down), "outages": len(down),
            "per_circuit_down_min": per_min, "intervals": down}

```

***Listing 15.2.** A simulated year: each circuit’s outages, the common-mode events, and the minutes during which every circuit was down at once. code/firm/slamodel/firm_slamodel.py*

![Simulation: minutes down in a year by percentile of 1 000 simulated years (zero drawn at 0.01). Two circuits in one duct have more years without an outage than one circuit, and worse bad years: a duct event takes twelve hours on average. Diverse circuits almost never go down together. Data: fig_buy.py.](https://one-course.com/images/onecourse/chapters/quant-14/nw-buying-connectivity/fig-ec6182d98613.svg)

***Figure 15.4.** Simulation: minutes down in a year by percentile of 1 000 simulated years (zero drawn at 0.01). Two circuits in one duct have more years without an outage than one circuit, and worse bad years: a duct event takes twelve hours on average. Diverse circuits almost never go down together. Data: `fig_buy.py`.*

What do the [service credits](#def-nw-buying-connectivity-sla) pay for these outages? Applying the published schedule of [Box 15.2](#dat-nw-buying-connectivity-sla) to each simulated month, the single circuit, on the single-connection terms, earns an average of 5.50 USD a year: a four-hour outage leaves the month at 99.44%, well above 95%. The two circuits in one duct, on the non-redundant terms, earn about 2 560 USD a year; the diverse pair, on the redundant terms, nothing. Against losses of the order of hundreds of thousands of dollars in the model, credits are not insurance: they are a signal of what the provider expects to deliver, and the firm insures itself by design.

## 15.5 Orders of magnitude for price

The published prices put a floor under a budget. A firm with two diverse 10-gigabit connections to the NYSE group pays about 320 000 USD a year for them in the model; the same firm’s two ultra-low-latency connections to an options exchange cost 360 000 USD (chapter 9); a radio route of 200 megabits between two New Jersey buildings costs more than half a million a year. What is not published (extranet services, [metro circuits](#def-nw-buying-connectivity-metro) between private buildings, long-haul wavelengths, managed racks) is priced by quotation, and chapter 29 treats those rows as estimates with ranges, never as facts.

**Method 15.8 (Buying a circuit).**

1. Say what the circuit is for: which strategies, what latency, what bandwidth, what happens when it fails.
2. Ask each provider for the path (ducts, streets, building entries, rooms), the one-way rack-to-rack latency and how it is measured, and its outage history.
3. Buy diversity, not duplicates: check that the second circuit shares nothing with the first, down to the building entry and the power.
4. Read the SLA for what it measures (monthly uptime, repair time), the credits and the exclusions; compare the credits with your own loss per minute.
5. Put every price, term and renewal date in the budget with its source (chapter 29), and re-check before renewal.

## 15.6 Tutorial: a year of outages

**Goal.** Price three designs from published fees, simulate their outages, and compare credits with losses. **End state:** [Figure 15.4](#fig-nw-buying-connectivity-ecdf) and Tables [15.1](#tab-nw-buying-connectivity-prices) and [15.2](#tab-nw-buying-connectivity-designs).

1. **The catalogue.** `firm_slamodel.CATALOGUE` holds published prices with their ledger rows; `SCHEDULES` the published credit tiers.
2. **The designs.** `nw_buy.designs()` builds one circuit, two in one duct and two diverse ones from `ASSUME` .
3. **Steady state and simulation.** `design_availability` ( [Listing 15.1](#lst-nw-buying-connectivity-avail) ) and `simulate_year` ( [Listing 15.2](#lst-nw-buying-connectivity-sim) ); `results()` fills [Table 15.2](#tab-nw-buying-connectivity-designs) .
4. **Credits.** `monthly_credits` cuts each simulated year into months and applies the schedule.

**What to change next.** Make the duct events rarer but longer; put the two diverse circuits with the same provider and add a shared provider fault; change the loss per minute to your own strategy’s figure.

## 15.7 Build: the SLA model

**Purpose.** Circuits as priced and sourced catalogue rows, their availability alone and in designs, the credits a contract pays and the losses it does not: the input of chapters 28 (resilience) and 29 (the plan).

**Interface.** `firm_slamodel`: `Tier`, `Circuit`, `Design`, `CATALOGUE`, `SCHEDULES`, `availability`, `design_availability`, `simulate_year`, `monthly_down_min`, `uptime_pct`, `credit`, `downtime_cost`, `expected_down_min`, `nines`.

**Rules.** Prices come only from the catalogue, with their source and date; failure rates are stated assumptions; a design’s common-mode events are explicit, never zero by default when circuits share anything.

**Acceptance tests.** `code/firm/slamodel/tests/`: availability by hand for the three designs, simulation against the steady state, monthly bucketing, and the credit tiers.

**Stretch.** Repair times that depend on the failure (a cut against a card), maintenance windows announced in advance, and a provider’s own common mode across its customers.

Sources and further reading

- AWS Direct Connect service-level agreement (July 2026).
- Nasdaq Data News #2017-8 (extranet providers); NYSE group connectivity fee schedule (April 2025).
- Chapters 9 to 14 for the venues’ and vendors’ own descriptions of their connections.

## 15.8 Exercises

**Exercise 15.1 ★.**

A circuit fails twice a year and takes four hours to repair each time. What is its availability, and how many minutes is it down a year?

**Solution of Exercise 15.1.**

MTBF $8\,766/2 = 4383\,\mathrm{h}$, MTTR $4\,\mathrm{h}$: availability $4\,383/4\,387 = 99.9088\%$, down $8\,766 \times 60 \times 0.000912 = 480$ minutes a year (eight hours).

**Exercise 15.2 ★.**

From [Table 15.1](#tab-nw-buying-connectivity-prices), what does a megabit of wireless bandwidth between Mahwah and Secaucus cost a month at 10 and at 200 megabits?

**Solution of Exercise 15.2.**

$9\,000/10 = 900$ USD per megabit a month at 10 megabits; $44\,000/200 = 220$ USD at 200.

**Exercise 15.3 ★.**

Under the single-connection schedule of [Box 15.2](#dat-nw-buying-connectivity-sla), how long must a connection be down in a 30-day month before any credit is due?

**Solution of Exercise 15.3.**

Below 95% of $30 \times 1\,440 = 43\,200$ minutes: more than 2 160 minutes, 36 hours, in the month.

**Exercise 15.4 ★★.**

Using [Proposition 15.7](#prop-nw-buying-connectivity-two), why do two circuits in one duct gain so little over one?

**Solution of Exercise 15.4.**

The pair is unavailable $u^2 + c$: with $u = 0.00091$ the independent term is $8.3 \times 10^{-7}$, negligible, and the duct’s $c = 6/8\,766 = 0.00068$ is of the same order as one circuit’s $u$. The second circuit removes the independent failures, which were never the problem, and keeps the common one.

**Exercise 15.5 ★★.**

List five things two “diverse” circuits can still share.

**Solution of Exercise 15.5.**

The building entry or the street near it; the [meet-me room](https://one-course.com/books/quant/14/en/chapter/9-colocation-products-and-how-they-are-sold#def-nw-colocation-products-and-how-they-are-sold-mmr) or the carrier cage; a patch panel or a switch at either end; the power supply of the racks; the provider itself (a software fault, a maintenance error, a bankruptcy) if both come from one; a wholesale fibre carried under two brands.

**Exercise 15.6 ★★.**

When is a [financial extranet](#def-nw-buying-connectivity-extranet) the right way to reach a venue, and when is it not?

**Solution of Exercise 15.6.**

When the firm needs many venues and data sources at moderate latency, or fast access to a new venue, and does not want to build and operate circuits to each. Not when the strategy’s edge is latency to that venue: the extranet’s own network and shared capacity sit in the path, and the firm cannot control them.

**Exercise 15.7 ★★★.**

*Coding.* With `firm.slamodel`, find the common-mode rate at which two circuits in one duct are no better than one circuit.

**Solution of Exercise 15.7.**

Solving $u^2 + c - u^2 c = u$ with the model’s circuits gives $c = 0.000911$, that is 0.67 duct events a year of twelve hours: at that rate the second circuit in the same duct buys nothing at all.

**Exercise 15.8 ★★★.**

*Find the flaw.* “Our provider guarantees 99.99% and pays credits, so an outage costs us nothing.”

**Solution of Exercise 15.8.**

The credit is a fraction of the month’s charge for the failed service (10% at the first threshold), credited against future bills, and it is the sole remedy; the firm’s trading loss during the outage is not covered. And 99.99% of a month still allows four minutes of outage a month without any credit at all.

## 15.9 Problem: Two Circuits Are Not Twice as Good

**Problem 15.1.**

Weekend problem — redundancy that is not

A firm connects to the NYSE group over 10-gigabit connections. Each fails twice a year and takes four hours on average to repair. Its two circuits run in one duct, which suffers half a common-mode event a year with twelve hours of repair; a diverse pair would cost 15% more. The firm loses 2 000 USD for each minute without a connection.

**Part I — One circuit.**

1. What are the circuit’s availability and expected minutes down a year?
2. What does it cost a year with its [cross-connect](https://one-course.com/books/quant/14/en/chapter/9-colocation-products-and-how-they-are-sold#def-nw-colocation-products-and-how-they-are-sold-mmr) ?
3. What is the expected annual loss?
4. What credit does the single-connection schedule pay for a four-hour outage?

**Part II — Two circuits in one duct.**

5. What are the pair’s availability and expected minutes down?
6. How much of that is the duct?
7. What does the pair cost, and what loss does it avoid compared with one circuit?
8. What does a bad year look like (95th percentile)?

**Part III — Two diverse circuits.**

9. What are the pair’s availability and expected minutes down?
10. What does diversity cost over the duct, and what does it save?
11. What credits do the three designs earn a year on average?
12. What would make the diverse pair fail together anyway?

**Part IV — The verdict.**

13. State the *named result* : the availability of two circuits sharing a duct against two diverse ones, their expected annual minutes down, and the [service credit](#def-nw-buying-connectivity-sla) against the trading loss.
14. Which assumption moves the answer most?
15. How would you verify a provider’s claim of diversity?
16. What should the SLA measure for a trading firm, beyond monthly uptime?
17. How do the firm’s systems use two circuits: active and standby, or both active?
18. Where does this analysis go in the connectivity plan of chapter 29?
19. What does the second circuit buy if it shares the duct?
20. In one sentence: what is redundancy?

**Solution of Problem 15.1.**

**Part I.**

1. 99.9088%, 479.6 minutes a year.
2. $12 \times (11\,000 + 600) = 139\,200$ USD a year.
3. $479.6 \times 2\,000 \approx 959\,000$ USD.
4. Nothing: four hours in a month leaves 99.44% uptime, above the 95% threshold.

**Part II.**

1. 99.9315%, 360.4 minutes a year.
2. 360.0 of the 360.4 minutes: half an event a year of twelve hours.
3. 278 400 USD a year; it avoids about 238 000 USD of expected loss against one circuit, less than its extra 139 200 plus the risk it keeps.
4. 1 992 minutes, 33 hours: worse than one circuit’s 1 541, because a duct event lasts twelve hours.

**Part III.**

1. 99.99992%, 0.44 minutes a year.
2. 41 760 USD a year more than the duct; it saves 720 000 USD of expected loss.
3. One circuit about 5.50 USD, the duct pair about 2 560 USD, the diverse pair nothing, under the schedules of their tiers.
4. A shared element missed in the survey (building entry, power, the provider’s own systems), a regional disaster, or a failure at the venue’s end, which no circuit fixes (chapter 28).

**Part IV.**

1. *Named result* : two circuits in one duct are available 99.93% of the time (360 minutes down a year, 721 000 USD of expected loss); two diverse circuits 99.99992% (0.4 minutes); the [service credits](#def-nw-buying-connectivity-sla) the duct pair earns are about 2 560 USD a year, less than 0.4% of its expected loss.
2. The rate and duration of common-mode events: the whole difference between the designs lives there.
3. Ask for the route maps of both circuits down to the street and the building entry, check them against each other and against the providers’ shared infrastructure, and test by failing one circuit.
4. Latency and its stability (not only availability), time to repair, advance notice of maintenance, and the provider’s diversity commitments.
5. Both active is better: traffic is spread or duplicated, failure detection is immediate, and the standby is known to work; active-standby needs regular fail-over tests to be trusted.
6. In its resilience section: each connection’s design, availability and cost, and the loss it protects.
7. Protection against the independent failures, which were a small part of the risk; it does not protect against the duct.
8. Two things that fail separately.

## 15.10 Interview questions

**Interview question 15.1 ★ developer.**

What is an SLA, and what does a [service credit](#def-nw-buying-connectivity-sla) compensate?

**Solution of Interview question 15.1.**

The contract’s statement of performance (availability, sometimes latency and repair time), how it is measured, and what the provider owes when it fails. A credit compensates the fee for the service not delivered, usually a fraction of a month’s charge; it does not compensate the customer’s losses.

*What the interviewer is looking for: Measurement period and method; credits as a fraction of fees; sole remedy; exclusions.*

**Interview question 15.2 ★★ developer.**

You have two circuits to the same venue from two providers. How can they still fail together?

**Solution of Interview question 15.2.**

Through anything they share: a duct, a building entry, a room, a panel, power, a wholesale carrier under both, or the venue’s side of the connection; and through common causes: a regional disaster, a configuration pushed to both.

*What the interviewer is looking for: Physical and logical common modes; verification by route maps and tests.*

**Interview question 15.3 ★★ developer, trader.**

When would you use a [financial extranet](#def-nw-buying-connectivity-extranet) instead of a direct connection to a venue?

**Solution of Interview question 15.3.**

For breadth and speed of set-up (many venues and feeds through one connection, a new venue in weeks) when the strategy is not latency-critical at that venue; not for the venues where the firm races.

*What the interviewer is looking for: Latency against convenience; the provider’s shared network; venue agreements and fees.*

**Interview question 15.4 ★★ developer.**

How do MTBF and MTTR combine into availability, and what does a second independent circuit do to it?

**Solution of Interview question 15.4.**

Availability is $\mathrm{MTBF}/(\mathrm{MTBF}+\mathrm{MTTR})$; with unavailability $u$ each, two independent circuits are down together $u^2$ of the time, for example $0.001^2 = 10^{-6}$; but any common mode $c$ adds directly and usually dominates.

*What the interviewer is looking for: The formula; squares for independence; common mode as the real risk.*

**Interview question 15.5 ★★ trader.**

Radio bandwidth costs hundreds of times more per megabit than fibre. Why do firms buy it?

**Solution of Interview question 15.5.**

For time, not bandwidth: a few megabits that arrive hundreds of microseconds or milliseconds earlier carry the messages that decide the races, and their value exceeds the price; everything else stays on fibre.

*What the interviewer is looking for: Value of latency; small messages; fibre back-up.*

**Interview question 15.6 ★★★ developer, researcher.**

Design the connectivity of a new trading office to three venues in two cities for 99.999% availability. What do you buy and what do you verify?

**Solution of Interview question 15.6.**

Two diverse paths from the office to each city (different providers, streets, entries, power), dual connections at each venue from two diverse points, both active; verify route maps, test fail-over regularly, monitor latency and loss on every path, and compute the design’s availability with explicit common-mode assumptions; five nines is about five minutes a year, which only diversity gives.

*What the interviewer is looking for: Diversity end to end; active-active; testing; the arithmetic of 99.999%.*