Back to blog

Architecture  ·  Design  ·  July 2026

Network Architecture
Designing for Failure

Resilience isn't a product you buy — it's a question you ask at every layer: "what happens when this dies?" Redundancy added randomly is expensive theatre; redundancy added where the single points of failure actually are is engineering. This is the SPOF audit — a systematic walk through every layer to find what takes you down, and what's worth fixing.

Architecture Resilience High Availability Redundancy SPOF

A single point of failure is any component whose death takes down more than you can accept. Every network has them; the question is whether you chose which ones to keep. Resilience engineering is not "add redundancy everywhere" — that's how you double the cost and the complexity while still missing the one SPOF that actually bites. It's a deliberate audit: walk each layer, name what happens when each thing fails, price the fix against the cost of the outage, and redundant-ify only where the math works.

The method is a layer-by-layer walk — power, links, devices, WAN, protocols, and services — asking the same question at each: if this single thing dies, what goes down, for how long, and is that acceptable? Where it isn't, you already have the tools from this series: FHRP for gateways, LACP for links, dual-WAN for uplinks, redundant cores for devices. This post is the framework that tells you where to spend them.

Reads alongside

01

The Audit — One Question, Every Layer

Walk the stack from the wall socket up. At each layer, name the single thing whose failure hurts, and decide if it's acceptable.

Layer / componentThe fix (if the outage is unacceptable)
Power (one PSU, one feed)Dual PSUs on different circuits/UPS; generator for critical sites
A single uplink cableLACP bond across two links (and two line cards)
The core switch/routerRedundant pair + FHRP for the gateway
The default gateway IPHSRP/VRRP virtual gateway
The single ISP / WAN linkDual-WAN (second provider, ideally different medium)
A routing/VPN pathBFD + a feasible backup path
DHCP / DNS / RADIUS serverA second server; anycast or failover pairing
The one engineer who knows itDocumentation & backups — the most-ignored SPOF

02

Match Redundancy to the Cost of Downtime

The same SPOF is worth fixing at one site and ignoring at another. What decides it is the cost of the outage, not the technical satisfaction of the fix.

A dead core switch at a 3-person office is a bad afternoon; at a 200-bed hospital it's a safety incident. Redundancy has real costs — the second device, the added config complexity, the new failure modes redundancy itself introduces (a split-brain FHRP, a misconfigured bond). So you price it: expected outage cost (frequency × duration × business impact) against redundancy cost (hardware + complexity + the risk the redundancy misbehaves). Fix the SPOFs where downtime is expensive or dangerous; consciously accept the ones where a short outage is cheaper than eliminating it. Writing down "we accept X as a SPOF because Y" is a legitimate, senior engineering decision — far better than pretending it doesn't exist.

⚠ Gotcha — Redundancy You Never Test Is a Belief, Not a Fact

The cruelest failures are the ones where the redundancy was there and didn't work: the backup ISP whose config drifted, the secondary firewall that was never in sync, the UPS with a dead battery, the failover that had never actually been triggered. Untested redundancy is worse than none, because it buys false confidence. Design the failover, then test it by causing the failure — pull the cable, kill the primary, drain the UPS — on a schedule. The broken-network and field-notes posts exist because failure drills are where you learn what your design actually does, versus what you assumed.

03

The SPOFs Everyone Forgets Reference

Teams that redundant-ify the obvious hardware still get taken down by these. They're cheap to fix and routinely ignored.

Forgotten SPOFWhy it bites
Both "redundant" links in one conduitOne backhoe cuts both — diverse physical paths matter
Both uplinks to the same ISP / same POPThe provider's outage is still your outage — different providers
Two power supplies on one circuitDual PSU, single breaker = single point of failure
No config backups / no OOB accessHardware survives; recovery doesn't. See OOB
One person holds all the knowledgeDocumentation is redundancy for the human layer
Expired certificate / licenceA silent time-bomb SPOF — monitor expiry dates

Takeaways

  1. Resilience is an audit, not a purchase. Walk every layer asking "what happens when this dies?" and fix deliberately.
  2. Redundancy everywhere is theatre — it doubles cost and complexity and still misses the real SPOF. Target the ones that hurt.
  3. Price it: expected outage cost vs redundancy cost (including the new failure modes redundancy adds). Consciously accept the SPOFs that aren't worth fixing.
  4. Untested redundancy is a belief, not a fact — cause the failure on a schedule, or you're just hoping.
  5. Diverse paths, diverse providers, diverse power — "redundant" links in one conduit or to one ISP aren't redundant.
  6. Documentation, backups, and OOB are redundancy too — the human and recovery layers are the most-ignored SPOFs.

Confident your "redundant" network actually survives a failure?

NOCTIS audits and designs for resilience across Crete — a real SPOF review across power, links, devices, and WAN, with redundancy where it counts and failover we prove by pulling the cable.

Book a Discovery Call →