how to build least efficient datacenter

Wetin you dey build

Every guide for this site teach you how to do things correctly: the commands for order, wetin correct result look like, and how to identify failure modes. This guide different. Today, just for play, we wan design the least efficient datacenter wey money, electricity, and pride fit produce.

We need one metric, so we go borrow the industry one: PUE, Power Usage Effectiveness — total facility power divide by the power wey actually reach computing equipment. A hyperscale datacenter dey run around 1.1: almost every watt dey do useful work. A decent enterprise server room dey manage 1.5. Our target na 4.0 or higher, wey mean say for every watt of computing, three more watts dey die for nothing. We go refer to this number many times, just as serious guides dey refer to backups.

Site selection: heat na the point

Cooling na the single largest overhead for real datacenter, na why ours go fight thermodynamics for its own home ground. The ideal location na attic. South-facing. Ideally with a skylight wey position to shine directly for the server, so that the machine go receive its own waste heat and the sun heat, collaboration between your electricity bill and a star.

For winter, cooling dey happen by opening the window. Real datacenters dey use outside air — the technique na free cooling, and dem dey engineer, filter, and control humidity for am. We go use am by accident, through a window wey dey allow rain, pollen, and at least one confused bird every quarter.

For true artistry, install an air conditioner, then place a space heater two feet from its thermostat, set two degrees warmer than the air conditioner target. Both machines go run continuously, forever, in perfect disagreement. The power company go send you card for Christmas.

One server, large, beloved

Redundancy dey dilute commitment. Our datacenter get exactly one server, and e big, because one single machine with 512 GB of RAM feel like infrastructure, whereas four small ones feel like a to-do list.

The server get name. No be hostname — a name. Gandalf, usually, or Odin. You no fit decommission Odin. Odin don dey up for five years:

$ uptime
 09:14:02 up 1847 days,  3:22,  1 user,  load average: 6.41, 6.38, 6.40

That number na point of pride, na why you go screenshot am and post am, and na why every attacker wey see the screenshot go see am as impressive: 1,847 days of uptime means 1,847 days of kernel vulnerabilities, wey nobody patch. Rebooting no dey option anyway — reboot na how you go discover which services dem start by hand for 2021 and dem never write for systemd unit. Nobody remember which ones. The server don become load-bearing for the organizational chart.

Storage: speed, and other ways to lose data

The disks set up for RAID 0, for performance. The zero refer to the number of disks wey fit fail. For maximum effect, stripe the array across storage wey get mixed provenance: two proper SSDs, one aging spinner, and a USB stick from one conference. The array reliable exactly as the conference stick, wey na the design.

Backups dey happen through one directory for the same array wey dem name backup_final_v2_REAL, wey contain one tarball of the previous naming scheme. Off-site backups dey represent by one sticky note wey write "set up off-site backups," wey, technically, dey stored off-site when you carry am home for your laptop lid.

A correct result look like: df reporting 97% usage, and one plan to deal with am next sprint.

Networking: a single strand of everything

The DNS server dey run for the machine itself, so that when the server fall, e go carry the DNS record wey you go use find out why. Dem dey call am consolidation.

The firewall dem disable am for 2021 — temporarily, to debug something. The debugging finish; the firewall no return. Every port for the router forwarded to the server "to save time later," and the router admin panel dey reachable from the WAN side with its factory password, for convenient remote management. Yours, and others'.

The server don dey run unusually warm lately, even by attic standards, and top show say the busiest process na something wey dem call xmrig. We assume say na the monitoring tool wey we dey use. We no install am — e appear by itself shortly after dem forward the ports, wey we take as sign say the ecosystem dey thrive. E dey monitor around the clock.

Power dey come through chain of consumer power strips wey combined length pass walking distance to the breaker panel — wey na efficient, in one sense, because you go dey visit the breaker panel many times.

Redundancy through complexity

After we refuse redundancy for where e matter, we now add am for where e no matter. The company homepage — one static HTML file — dey serve by a twelve-node Kubernetes cluster. This achieve wetin engineers dey call resume-driven architecture: the page load for the same forty milliseconds wey nginx for don deliver, but e fit now fail for ways wey go require a consultant.

For isolation, the cluster itself dey run inside a virtual machine inside a virtual machine inside a virtual machine, each layer dey add security the way each layer of a turducken dey add bird. The contact form na nine microservices. Two of dem never get invoked. One of dem na load-bearing and nobody know which one.

Heating as a service

Modern server dey convert electricity into computation and heat, and we wan maximize the second output. A media server with no GPU na the classic play: CPU-transcoding one single 4K stream go pin sixteen cores and warm one small bedroom, a space heater wey dey also play movies. The ambitious operator graduate to running a large language model on CPU — a 70-billion-parameter space heater wey get API, dey produce tokens at a rate wey best to measure seasonally.

The monitor watches itself

Observability matter, so we deploy a self-hosted uptime monitor — for the same server wey e dey monitor. When Odin die, the monitor die with am, and here na the elegant part: no alerts fire. No alerts means no incidents. No incidents means perfect uptime, as it measure. The monthly report never look better.

Alert emails, for completeness, dey relay through a mail server wey also dey run for Odin. The alerting pipeline so, fully self-contained, the way snake wey dey eat its own tail dey fully fed.

The uncomfortable part

Na this section I don dey put for back. None of this na fiction. The beloved irreplaceable server, the RAID 0 with backups for the same volume, the firewall wey dem disable "temporarily," the Kubernetes cluster wey dey serve one page, the monitor wey dey watch itself — I don see every one of these for production. Some of dem I don see this year. One or two of dem, for my early days, I build.

Wetin real efficiency look like na boring thing, na why e lose the argument for the moment but e win am over a decade: a PUE wey you no dey think about because someone else engineer am. Machines sized to their workload instead of to their owner's self-image. A blast radius, wey dem consider before explosion. Backups wey dem test by restoring dem, on a schedule, with a calendar reminder and no heroism. Redundancy wey dey dull — two of the cheap thing dey beat one of the magnificent thing, every time, for every failure wey I don ever get paged for.

And the most efficient datacenter wey you fit run na the one wey you no dey run. A VPS hand the power, cooling, redundancy, and 3 a.m. hardware failures to people wey dey do dem at scale, boringly, wey na the highest compliment infrastructure fit earn — and e leave you the genuinely fun part, wey na running your own services on top of it, on a machine wey you fit afford to lose, wey na the only kind wey you should ever experiment on.

FAQ

I suppose I should actually do any of this?

No. Every section for this guide na documented anti-pattern wey don kill many weekends. If your current setup resemble more than two sections, skip to the last question for this FAQ — in the order dem give am, because the order na the triage.

Wetin be good PUE, actually?

Hyperscale datacenters dey run around 1.1, a well-run enterprise room dey manage 1.4 to 1.6, and an uncooled closet with a space heater feud fit genuinely exceed 3. You no fit meaningfully compete with 1.1 for house, wey na the quiet economic argument for renting compute from person wey fit do am.

Na real thing to heat building with servers?

Yes — if dem do am properly. District-heating projects for several countries dey capture datacenter waste heat through heat exchangers and dem dey pipe am into homes, by design, with engineering and contracts. The satire above no be say server heat fit warm room; na say dem dey do am by accident and dem dey call the accident a strategy.

My server already look like this. Wetin I go do first?

Backups, tonight, to somewhere wey no be the server, and then a test restore — untested backup na just rumor. Second, patches and the reboot wey you don dey avoid, for one planned window, so you go learn wetin go break while you dey watch. Third, split the single point of failure: move DNS and monitoring off the box. Everything else fit wait for one calmer week; those three no fit wait.