[Beginner] · The building. Power, heat, redundancy tiers, and why “availability zone” is a physical fact rather than a marketing word.
A server has to live somewhere. Under a desk works until it doesn’t — until the cleaner unplugs it, the room hits 40°C in August, or the building loses power for six hours. A data centre is a building whose entire purpose is to remove those failure modes, and it is more interesting than it sounds, because its constraints leak upward into things you’ll actually configure.
What the building is actually for
Four things, in rough order of how much they cost:
- Power that doesn’t stop — mains, plus UPS batteries for the seconds-to-minutes gap, plus diesel generators for the hours-to-days gap
- Heat removal — a fully loaded rack dissipates 5–20 kW as heat, and a hall has hundreds of racks
- Connectivity — multiple physical fibre routes from multiple carriers, entering the building at different points
- Physical security — mantraps, biometrics, cameras, and a real answer to “who touched that machine”
Compute is the cheap part. The building, the power contract and the cooling dominate the cost, which is exactly why the cloud’s economics work: those costs amortise brutally well across scale.
Racks and the unit that everything is measured in
The 19-inch rack is the organising standard, and it is remarkably old — it descends from railway signalling equipment. Vertical space is measured in U (1U = 44.45 mm). A full rack is 42U.
A rack contains servers, a top-of-rack switch (see the DC network), power distribution units, and cable management. Racks stand in rows; rows form halls.
Power density is the real limit, not space. A rack might have room for 42 1U servers but only enough power and cooling budget for 20. Empty U in a modern hall usually means the power circuit is full, not that nobody wants the space.
Hot aisle / cold aisle
The single most important physical arrangement, and once you see it you can’t unsee it.
Every server pulls cool air in the front and exhausts hot air out the back. So racks are arranged so that fronts face fronts and backs face backs:
COLD AISLE HOT AISLE COLD AISLE
↓ ↓ ↓ ↑ ↑ ↑ ↓ ↓ ↓
┌────────┐┌────────┐ ┌────────┐┌────────┐
│ rack → ││ ← rack │ │ rack → ││ ← rack │
└────────┘└────────┘ └────────┘└────────┘
front back back front front back
Cool air is delivered to the cold aisles (often through a raised floor with perforated tiles), and hot exhaust is collected from the hot aisles and returned to the cooling plant. Modern halls contain one aisle or the other with physical barriers so the two air masses never mix.
The failure this prevents: if hot exhaust recirculates into an intake, that server runs hot, throttles, and eventually shuts down — while the room’s average temperature reads fine. Mixing air is the enemy; the entire discipline is about keeping two air masses apart.
The efficiency metric is PUE (Power Usage Effectiveness) — total facility power ÷ IT equipment power. PUE 2.0 means you burn a watt on cooling and overhead for every watt of compute. A good modern facility is 1.1–1.2. Hyperscalers report figures near 1.1, which is a large part of why renting is cheaper than a server room in your office, where a PUE of 2 would be optimistic.
Redundancy, and the N+1 notation
Redundancy is described by how many spare units you have relative to what you need:
- N — exactly enough capacity. No spare. One failure is an outage
- N+1 — one spare component. Any single unit can fail or be serviced
- 2N — a complete duplicate of everything. Two independent power paths
- 2N+1 — duplicate, plus a spare in each
The Uptime Institute tier ratings package this into four levels:
| Tier | Shape | Concurrently maintainable? | Rough availability |
|---|---|---|---|
| I | Single path, no redundancy | No | ~99.67% (~29 h/yr) |
| II | Single path, redundant components | No | ~99.75% (~22 h/yr) |
| III | Multiple paths, one active | Yes — service anything without downtime | ~99.98% (~1.6 h/yr) |
| IV | Multiple active paths, fault tolerant | Yes, and survives a failure during maintenance | ~99.995% (~26 min/yr) |
Tier III is the commercially interesting line, and the reason is maintenance rather than failure. Below it, routine work on the power path requires taking the load down. Most colocation you’d actually buy is Tier III.
Read those availability figures carefully. Tier IV’s 99.995% is the facility. Your application’s availability is that number multiplied by everything else in the chain — your code, your database, your deploy process. The building is rarely your weakest link, which is the useful thing to know. See availability and reliability for how these numbers compose.
Where your servers actually sit
Four models, and the difference is who owns which layer:
| Model | You own | They own |
|---|---|---|
| On-premises | Everything, including the building | — |
| Colocation | The servers | Space, power, cooling, connectivity |
| Managed hosting | The application | The hardware too |
| Cloud | A workload | All of it |
Colocation is the one people forget exists, and it’s still the right answer for predictable, steady, high-volume workloads where cloud pricing stops being a bargain. The cloud’s advantage is elasticity and speed, not unit cost — several well-known companies have repatriated steady workloads and saved substantially. The trade is that you get your lead times and your capacity planning back.
Regions, availability zones, and what the words mean physically
This is the payoff, and it’s why the building is worth a note.
A cloud region is a geographic area — eu-west-1 (Ireland), us-east-1 (Northern Virginia). Regions are genuinely separate: separate power grids, separate staff, and services are usually region-scoped. Data doesn’t move between them unless you move it, which is what makes regions the unit of data residency and compliance.
An availability zone is one or more physically distinct data centres within a region, with independent power, cooling and networking — but connected to the other AZs in the region by dedicated low-latency fibre, typically single-digit milliseconds round trip.
That number is the whole design:
- Far enough apart that one flood, fire, or substation failure cannot take out two
- Close enough that synchronous replication between them is practical
So the standard advice — spread across at least two AZs — is not a cloud-vendor ritual. It’s the statement that an AZ is the blast radius of a physical event, and one AZ is a single building’s worth of risk no matter how many virtual machines you run inside it.
The corollary is worth stating plainly: three instances in one AZ are not a redundant system. They share a roof, a power feed, and a cooling plant. Multi-AZ protects against the building. Multi-region protects against the region — at the cost of latency and, usually, of strong consistency. See distributed systems for what that trade costs.
Related
- the data centre network — how the racks are wired together
- servers — what fills the racks
- cloud fundamentals — regions and AZs as an API
- availability and reliability — turning these numbers into design
- congestion control — why DC networks behave unlike the internet
Source: [reference] — from the freeCodeCamp IT Fundamentals course (module 3), extended with Uptime Institute tier definitions and cloud-provider AZ documentation.
Software engineering
Infrastructure & security
Computer science
Hardware & engineering