Uptime SLAs That Mean Something

An uptime SLA looks like a promise and behaves like a refund policy. Suppliers know this. Customers frequently do not, and sign a service level agreement believing they have bought availability when what they have bought is a small discount in the event they do not get it.

That is not necessarily a bad deal. It is a different deal from the one most people think they are making, and the difference matters when the system is down and somebody asks what the contract says.

Three nines sounds close to perfect and allows 43 minutes of downtime a month. Four nines allows four. If your business can absorb a 43 minute outage during a working afternoon, 99.9% is fine and you should not pay for more. If it cannot, no number of nines helps, because the agreement pays you a credit rather than preventing the outage.

What an Uptime SLA Promises in Minutes

The percentages compress a large range into a small-looking difference. Written as time, the differences are obvious.

Two things follow. The step from 99.9% to 99.99% is a tenfold reduction in permitted downtime and typically a much larger increase in cost, because it requires redundancy that removes every single point of failure including the deployment process. And anything at or beyond five nines is below the time it takes a human to notice and respond, so it can only be met by systems that fail over without anybody involved.

Decide the figure from what an outage actually costs you per hour, not from what sounds impressive. Most business systems are honestly served by 99.9%.

Service Credits Are Not Compensation

The remedy in nearly every uptime SLA is a credit against future fees, usually a percentage of the monthly charge, and often capped at that month’s total. Google Cloud’s compute SLA is a representative example: credits only, claimed by the customer, capped.

That means the maximum you can recover is roughly what you paid for the period, which for most services is far less than an outage costs. A platform charged at a few hundred pounds a month cannot compensate a day of lost trading, and no supplier will agree to terms where it could.

Credits are also frequently claim-based rather than automatic. You must notice the breach, calculate it, and submit a claim inside a window, sometimes as short as thirty days. Suppliers are not obliged to volunteer that they missed the target, and most do not.

The practical conclusion is to stop treating the credit as protection. It is a signal of how seriously the supplier takes their own number. What protects you is architecture, and the recovery planning covered in our disaster recovery guide.

The Clauses That Decide Everything

What counts as downtime. Total unavailability, or degraded performance too? A service responding in thirty seconds is unusable and, under most definitions, available. Get a latency threshold into the definition if response time matters.

How it is measured, and by whom. Almost always by the supplier, using their own monitoring, from their own network. That is measurement of their infrastructure rather than of your experience. Agree the measurement point, and run independent monitoring regardless, because you cannot dispute a number you did not observe.

Exclusions. Planned maintenance is usually excluded entirely, which means a supplier can be unavailable for hours a month without breaching anything. Check how much notice is required and whether there is a cap. Also excluded: anything attributed to your configuration, to a third party, or to force majeure, and those categories can absorb a lot.

The measurement window. Monthly is standard and favours you. Annual measurement lets a supplier take a long outage and still meet the target, because a full day down is inside the yearly allowance at 99.9%.

Response versus resolution. Many agreements commit to responding within a period and say nothing about fixing. A one hour response and an unbounded fix is a common shape and is much weaker than it reads.

What to Ask For Instead

Push on the things that change behaviour rather than on the headline number.

A cap on planned maintenance, with notice and a window outside your peak hours. This often affects real availability more than the percentage does.

Status transparency, meaning a public status page and a commitment to post incidents promptly. A supplier who reports their own outages is telling you something about how they operate.

A defined escalation path with names or roles and times, so an incident does not begin with finding out who to contact.

Incident reports after significant outages, describing what happened and what changed as a result. This is the single most informative document a supplier can give you, and a reluctance to provide one is itself informative.

Termination rights on repeated failure. Chronic underperformance should let you leave without penalty. That right is worth more than any credit, because it is the only remedy proportionate to the harm.

If You Are the One Offering an SLA

Offer a number you can meet on your worst month, not your average one, and measure it yourself before you commit to it. Suppliers who commit to a figure they have never verified discover the gap in public.

Define planned maintenance honestly and keep to the window. Publish a status page. Write the incident report even when nobody asks, because the customers who stay are the ones who believe you know what happened.

And price the tiers by what they actually cost you to deliver. Four nines is not a marketing decision, it is redundant infrastructure, automated failover and a deployment process that cannot take the system down, and all of that has a bill attached. The operational groundwork in our CI/CD pipeline guide is where most of it starts.

Mecanik builds and operates systems under availability commitments as part of our server security work. The commitment worth making is nearly always the one you can hold during a bad week.

Frequently Asked Questions

How much downtime does 99.9% uptime allow? 8.76 hours a year, or 43.2 minutes a month. The step to 99.99% cuts that to 52.6 minutes a year and 4.32 minutes a month, a tenfold reduction that typically costs far more than ten times as much to deliver because it requires removing every single point of failure including the deployment process.

Are service credits real compensation for downtime? No. Credits are usually a percentage of the monthly fee and capped at that month’s charge, so the most you can recover is roughly what you paid, which is far less than an outage typically costs. They are also often claim-based with a short window, and suppliers rarely volunteer that they missed the target.

What should I check in an uptime SLA? What counts as downtime and whether degraded performance is included, who measures it and from where, what the exclusions cover, whether planned maintenance is capped and noticed, whether the window is monthly or annual, and whether the commitment is to respond or to resolve.

Why does the measurement window matter? Monthly measurement favours the customer. Annual measurement lets a supplier take one long outage and still meet the target, because a full day of downtime sits inside the yearly allowance at 99.9%. The same failure would breach a monthly agreement outright.

What is worth negotiating instead of a higher number? A cap on planned maintenance with notice outside peak hours, a public status page with prompt incident posting, a named escalation path with times, written incident reports after significant outages, and termination rights on repeated failure. The last is the only remedy proportionate to the harm.