Cloud & infrastructure
High availability: what “99.9%” actually means for your business
SLAs, single points of failure, power cuts and a staged architecture: building realistic availability for a business in Cameroon.
Contents (10 sections)
- What “99.9%” really allows as downtime
- High availability, backup, recovery: three answers to three problems
- Where are your single points of failure?
- Power first: power cuts, UPSs and solar
- Two Internet connections beat one big subscription
- Degraded mode: what does the team do during the outage?
- A staged progression
- Measure it: without a log, availability is an opinion
- Frequently asked questions
- Assessing your real availability
“Our system is 99.9% available” is a reassuring sentence, and a rarely understood one. Translated into real time, it permits 8 hours and 45 minutes of downtime a year. For internal email, that is acceptable. For a supermarket checkout in Bonamoussadi on a Saturday afternoon, it is a day of lost revenue.
High availability is not a product you buy: it is a trade-off between what downtime costs and what redundancy costs. And before talking about servers, start with a simpler question: what happens in your business when the power goes off?
What “99.9%” really allows as downtime
The downtime each level permits:
| Availability | Per year | Per month |
|---|---|---|
| 99% | 3 days 15 h | 7 h 18 min |
| 99.5% | 1 day 19 h | 3 h 39 min |
| 99.9% | 8 h 45 min | 43 min |
| 99.95% | 4 h 22 min | 21 min |
| 99.99% | 52 min | 4 min |
Two observations matter more than the table.
An SLA is not a service guarantee: it is a penalty clause. A host that misses its commitment grants a credit on a fraction of the monthly bill, which you have to claim yourself: at AWS, for example, 10% to 100% depending on how bad the outage was, as credit against future bills rather than cash. That never makes up for the trading day you lost.
Availabilities multiply. Take some assumptions, for the sake of the example: the application at 99.9%, the Internet link at 99%, the power supply at 98%. What your users actually experience drops to about 96.9%, more than eleven days of downtime a year. The chain is only as strong as its weakest link; replace these assumptions with your own measurements.
High availability, backup, recovery: three answers to three problems
High availability prevents interruption, through redundancy: two links, two power feeds, two servers.
Backup recovers data that was lost or corrupted. It prevents no downtime at all. See our guide to backup and disaster recovery.
Disaster recovery restarts elsewhere after a fire, a flood or a theft.
None replaces the others. A redundant cluster that faithfully replicates a corrupted database gives you a corrupted database, twice.
Where are your single points of failure?
A single point of failure is anything whose failure stops everything. Walk the chain and ask, at each layer: if this fails right now, what happens?
- Power — one meter, no UPS, no backup supply.
- Connectivity — one operator, one router, one cable.
- Hardware — one server, one disk without RAID, one access point.
- Data — one copy, or several copies in the same place.
- People — one person who knows the passwords and the procedure.
- Suppliers — one provider, unreachable at the weekend.
The human one is the most often overlooked, and often the most expensive.
Power first: power cuts, UPSs and solar
Before buying anything, measure. For one month, log every cut on the grid, run by SOCADEL, the state-owned company that replaced ENEO under a decree of 4 May 2026: the time, the duration, and what stopped. That log tells you whether your problem is counted in minutes or in hours, and the answer is not the same.
| Option | Covers | Limit |
|---|---|---|
| UPS | A few minutes, clean shutdown | Not built to last for hours |
| Surge protection | Violent power returns | No runtime at all |
| Hybrid inverter and batteries | Several hours, depending on batteries | Batteries to size and replace |
| Solar panels | Recharging batteries by day | Little or no output at night |
| Generator | Long outages | Fuel, maintenance, no instant start |
To size a UPS, add up the power in watts of what it must protect: server, Internet box, switch, till, receipt printer. Choose a model whose rating in watts, not only in VA, exceeds that total with a margin. Then test the real runtime by cutting the power, once a quarter.
A sensible order of priority: a UPS with clean shutdown, surge protection, extended runtime (batteries, solar or a generator), and only then IT redundancy.
Two Internet connections beat one big subscription
A second link from a different operator — Camtel, MTN or Orange, depending on your area — even a modest one, buys more real availability than doubling bandwidth with the same provider. Two links on one operator often share the same infrastructure, and therefore the same outage.
A router that fails over automatically from one link to the other saves you waiting for an employee to notice. Test the failover once a month: unplug the main link and check that the till keeps working.
Degraded mode: what does the team do during the outage?
For a shop, being able to take payments offline and synchronise when the network returns is worth more than a perfect architecture that is unavailable. The Odoo point of sale, for example, has an offline mode — see our guide to Odoo ERP in Cameroon.
For the rest, write the procedure down: who decides to switch to degraded mode, how sales are recorded (a numbered notebook, for instance), who re-enters them when the network returns, and who checks. The question is not only “how do we prevent the outage?” but “what does the team do during it?”
A staged progression
Aiming for 99.99% from the start is neither necessary nor sensible.
Stage 1 — remove the catastrophic. UPS, surge protection, RAID on the server, automated off-site backup, a written procedure.
Stage 2 — remove the waiting. A second Internet link, spare parts on site, degraded mode for checkout, monitoring with alerts, a support contract with defined response times. See Nematek One Support.
Stage 3 — remove the interruption. Redundant servers, automatic failover, replication to a second site, regular failover tests. See our cloud, DevOps and DevSecOps services.
Start with stages 1 and 2: they cost less and remove the longest outages. Stage 3 is justified when an hour of downtime costs more than its annual price.
For deeper design guidance, the Azure Well-Architected reliability documentation is very practical, and the Uptime Institute tier classification shows why redundancy is paid for in steps. NIST SP 800-34 covers the planning side.
Measure it: without a log, availability is an opinion
- Monitoring that tests the service from outside, not only from the local network.
- An alert that reaches someone genuinely reachable, with an escalation path if they do not answer.
- An incident log: date, duration, cause, fix, power cuts included. Three months of logging reveal the dominant cause better than any intuition.
Frequently asked questions
What availability should an SME target?
It depends on what an hour of downtime costs. Estimate it concretely: hourly revenue, idle staff, lost orders. If an hour costs little, 99% is enough; if it costs the equivalent of a day, 99.9% becomes worth paying for. The figure follows from the calculation; it is not chosen by ambition.
Does redundancy double the cost?
Rarely exactly: the second piece of equipment adds cost, but engineering and operations do not necessarily double. Complexity does increase, though, and poorly mastered redundancy can cause more outages than it prevents. Redundancy that has never been tested is a risk, not a protection.
Does the cloud make high availability automatic?
No. A cloud provider supplies available building blocks; the architecture is still yours to design. A single virtual machine at a large provider is no more available than a well-run on-site server, and it also depends on your Internet connection.
How do we test failover without risking production?
By planning it. Pick a low-activity window, warn the teams, run the failover, time it, then return to the original state. A failover that has never been tested is very likely to fail on the day you need it.
Do we need a generator or solar panels?
That depends on how long and how often your power fails, which your one-month log will tell you. A UPS covers minutes and allows a clean shutdown. Batteries, recharged from the grid or by solar, cover hours. A generator covers long outages, at the cost of fuel and maintenance.
Assessing your real availability
The useful starting point fits in two documents: an inventory of your single points of failure and a costed estimate of one hour of downtime. They are enough to rank your investments.
We run this assessment as part of our cloud and infrastructure, security cameras and networks and IT support services. Request a free assessment.
Update history
- Updated: ENEO has become SOCADEL (decree of 4 May 2026), an SLA pays out in credits (AWS as an example) and the NIST SP 800-34 link points to the version in force.
- Rewritten: ENEO power cuts, a UPS, hybrid inverter, solar and generator table, stated assumptions, a second Camtel, MTN or Orange link and a degraded mode.
- Rewritten as a practical guide: the power supply before server redundancy, the “nines” of availability and verified sources.
Only a change of substance moves the update date; a re-read with nothing to change moves only the “Checked” date. Something looks out of date? Tell us.
Sources
Sources checked
- 10% to 100% depending on how bad the outage was, as credit against future bills rather than cash aws.amazon.com
- replaced ENEO under a decree of 4 May 2026 crtv.cm
- Azure Well-Architected reliability learn.microsoft.com
- Uptime Institute tier classification uptimeinstitute.com
- NIST SP 800-34 csrc.nist.gov
Become a Nematek ambassador — free application and free training. You present Nematek One solutions and follow each customer through to the signed contract; your commission is paid once the solution is deployed and the customer has paid.