/notes / 11 Nov 2026

Why a 100% SLA isn't feasible in technology

A customer asked me why the cloud provider doesn't just replicate everything to another region and promise 100% uptime. The honest answer has more to do with physics, change and regulators than with greed.

· 6 min read

Cartoon of a grinning salesman holding a gold "100% SLA" sign while an excavator cuts a fibre cable, a cat chews a network cable, a regulator stamps "NOT HERE" and a tired engineer holds a "DR test?" mug.

Last week I hosted a customer event, and one of the guests asked the kind of question I secretly love.

His point, more or less: "You run the data centres. You have regions everywhere. Why don't you just replicate every service to another region, fail over automatically, and give us 100%? Why do I have to design my own DR, and then pay you extra for it?"

At first, we answered the question using logical answers; the SLA is what it is for a reason. Systems need patches, updates and fixes. Some downtime comes from things nobody controls, like the network between him and us. All true. But on the drive home I kept turning it over, because I don't think we gave him the strongest part of the answer. So here's the version I wish I'd given.

100% is a very specific number

People treat 100% like "very, very reliable". It isn't. It means zero. Not one minute of downtime, ever, for the life of the contract.

For comparison:

  • 99.9% allows about 8 hours 45 minutes of downtime a year
  • 99.99% allows about 52 minutes
  • 99.999% allows about 5 minutes

Every extra nine costs a lot more than the one before it, and the last few minutes are the most expensive minutes in all of IT. And 100% isn't one more nine. It's a promise that nothing will ever go wrong. Nobody who has spent a night on a bridge call would sign that, and if someone does, read the penalty clause. A 100% SLA usually means "we'll refund you 10% of this month's bill when it breaks", which is a pricing decision, not an engineering one.

Most outages come from change, not from disasters

When people picture DR they imagine a flood or a fire taking out a data centre. That happens, but it's rare. What takes systems down far more often is someone changing something: a patch, a config push, a certificate that expired, a new version that behaves slightly differently under real traffic.

You can't stop changing things. A system that never gets patched is a system that gets breached. So the people running it will keep touching it, and every touch carries a little risk.

Here's the part people miss: automatic replication copies the bad change too. If a corrupted record or a broken config goes into the primary, a perfectly working replication link delivers it to the secondary within seconds. Two regions, same problem. Replication protects you from losing a place. It doesn't protect you from your own mistakes.

Physics gets a vote

Say we did replicate everything to another region automatically. The next question is: synchronously or not?

Light in fibre covers roughly 200 km per millisecond. Two regions a few thousand kilometres apart means every single write waits tens of milliseconds for the far side to confirm. For a payment system doing thousands of commits a second, that's not a small tax. Applications slow down, timeouts start appearing, and you've made the system worse in order to protect it.

So you go asynchronous instead. Now the secondary is always a little behind, and if the primary disappears you lose whatever hadn't crossed yet. Maybe a second of data, maybe more.

Which one is right? It depends entirely on the workload. A trading platform and an internal wiki don't want the same answer. The provider doesn't know which of your systems can lose five seconds of data and which can't lose one transaction. You do.

And then there's the regulator

This is the bit I wish I'd said in the room.

If the provider decides on its own where your data fails over to, it has made a governance decision on your behalf. In a lot of industries that decision isn't the provider's to make, and sometimes it isn't even yours.

Think of a bank here in the UAE, or a hospital in Saudi Arabia, or a government entity anywhere in the region. Many of them operate under rules that say certain data must stay inside the country, or inside approved locations, or can only move with prior approval. A cheerful automatic failover to "the nearest healthy region" could put customer records in a jurisdiction the regulator never approved. The outage might be over in four minutes, but the compliance problem lasts a lot longer.

So who should pick the DR region? The organization that answers to the regulator. That's why the choice sits with the customer: which region, which services, what data class, how far behind it's allowed to be, and who presses the button. Look at the cartoon again. The guy with the clipboard is the reason the button isn't automatic.

"Fine, but why do I pay for it?"

Because a standby region isn't an idea. It's real compute, real storage and real network capacity sitting in another building, reserved for you, waiting for a bad day.

If every provider built that in for every customer by default, it wouldn't be free. It would just be hidden inside everyone's price, including the customers whose dev and test environments don't need a second region at all. Your finance team, and anyone doing FinOps properly, would much rather pay for DR where the business needs it than subsidize it everywhere.

Not every workload deserves the same protection. Some need a hot standby. Some are fine with a backup and a tested restore. Some can be down for a day and nobody notices. Paying for exactly the protection you chose is the fair version.

What I'd actually ask for instead

If I were that customer, I'd stop chasing 100% and ask three questions:

  1. For each critical system, how much downtime can the business really live with, and how much data can it lose?
  2. Where is my data allowed to live when things go wrong, and has my compliance team signed off on that?
  3. When did we last test the failover, end to end, with real people following the runbook?

The third one is the uncomfortable one. Plenty of companies have a DR design. Far fewer have a DR plan that has actually been run. A tested 99.95% beats an imagined 100% every time.

So no, nobody can honestly sell you 100%. What a good provider can give you is a clear number, the building blocks to go beyond it, and the freedom to decide where your data goes. The deciding part is yours, and in regulated industries, that's exactly how it should be.

And to the gentleman who pushed us so hard yesterday: thank you. It was the best question of the day.

These are my own views, not an official position of my employer.

clouddisaster-recoveryslagovernanceresilience

More notes