The machine, the building, and the wiring — the four notes this folder’s neighbours all assumed.

~6,300 words across 5 notes. Built August 2026. [reference].

The one idea: every abstraction above this — a container, an instance type, an availability zone, a VPC — is a rental agreement over a physical constraint. Knowing the constraint tells you where the abstraction leaks.

Why this exists

Cloud fundamentals opens with “the cloud is someone else’s computers” and then never mentions the computers. 01-linux starts with an OS already running on a machine that already exists. hardware/ covers the layer below software but means microcontrollers and PCBs, not server rooms.

So the vault had a gap exactly one layer wide: between a soldered circuit board and a rented t3.medium, nothing explained the actual machine, the building it lives in, or why an availability zone is a physical fact rather than a marketing word.

This is numbered 00 because it sits under Linux, not after it. It is also the shortest section in devops/, deliberately — it exists to make four ideas concrete, not to turn anyone into a facilities engineer.

Reading order

  1. Servers — [Beginner] — the same five components as your laptop, different economics; ECC, IOPS, and the four problems with one-app-per-server that produced everything else
  2. Virtualisation and Hypervisors — [Beginner → Intermediate] — Type 1 vs Type 2, oversubscription and steal time, and why containers are not smaller VMs
  3. Data Centres — [Beginner] — power, hot aisle/cold aisle, PUE, the Uptime tiers, and what an availability zone physically is
  4. The Data Centre Network — [Intermediate] — east–west traffic, leaf-spine, oversubscription ratios, VXLAN overlays, and incast
  5. Getting Into Infrastructure — [Beginner] — the roles and what separates them, hands-on-metal vs infrastructure-as-software, where certifications actually help, and what to build

The things worth carrying

  1. A server isn’t a different kind of machine — it’s the same parts, specified for concurrency and uptime. ECC is the giveaway → 01
  2. IOPS, not capacity, is the storage number that decides performance. Five orders of magnitude between an HDD and an NVMe on random reads → 01
  3. Underutilisation, not compute scarcity, is what made virtualisation inevitable. You can’t fix a 15%-utilised fleet with faster machines → 01
  4. CPU and memory virtualisation are nearly free; I/O is where the overhead lives. So look at disk and network first → 02
  5. Steal time is your bill for someone else’s busy hour. The visible face of oversubscription → 02
  6. A VM virtualises hardware; a container virtualises an OS. Which is why Docker on macOS is running a Linux VM → 02
  7. Once a running machine is a file, live migration exists — and hardware maintenance stops being an outage → 02
  8. Power density, not floor space, is what fills a data centre → 03
  9. An AZ is the blast radius of a physical event. Three instances in one AZ share a roof → 03
  10. The facility is almost never your weakest link. Tier IV is 26 minutes a year; your deploy process is worse → 03
  11. 70–80% of data centre traffic is east–west, which is why the three-tier tree gave way to leaf-spine → 04
  12. A VPC is an overlay, not a wire. Nothing physical changed when you chose 10.0.0.0/16 → 04
  13. SAN gives blocks, NAS gives files, S3 gives neither. EBS/EFS/S3 in one line → 04

Where this connects

08-cloud/The direct consumer. Instance types, regions, AZs — this is what they’re renting
03-docker/Note 02’s other half, in depth
networkingNote 04 is the DC-shaped inversion of that course’s assumptions
OS: isolationThe kernel primitives under containers
availabilityWhere the tier numbers turn into design
hardware/The layer below this one — boards, not buildings

The honest note

[reference], and the most reference-y section in devops/ — I have not racked a server or walked a data centre hall. Everything here is read, cross-checked, and arranged.

What would close the gap:

  1. Run a Type 2 hypervisor properly. Build a three-VM lab, deliberately oversubscribe it, and watch %st appear in top under load. Note 02 becomes an observation instead of a claim
  2. Measure the thing you rent. fio against an EBS volume vs an instance-store NVMe vs a local disk. The IOPS table in note 01 is only real once you’ve reproduced it
  3. Take one workload and place it badly on purpose — same-AZ vs cross-AZ vs cross-region — and record the latency difference. That’s note 03’s single-digit-millisecond claim, verified
  4. Trace a packet out of a container and count the encapsulation. tcpdump on the underlay while pinging across a VPC makes note 04’s overlay point undeniable
  5. Ask for a tour. Most colocation providers will show a prospective customer a hall. Twenty minutes in a cold aisle beats any amount of reading here

What’s missing: liquid and immersion cooling in any depth, power distribution below the rack PDU, Fibre Channel specifics, RDMA/RoCE beyond a mention, GPU-cluster fabrics (NVLink, InfiniBand) — a real gap given the GPU course — and anything about the economics of actually buying capacity.

→ Reading is not a rank.

↑ Back to DevOps