Service level agreements explained

People read "99.9% uptime SLA" as a promise that the service will be available 99.9% of the time. It is not. It is a statement that if availability falls below 99.9%, measured in the provider's defined way, excluding the provider's defined exclusions, and if you notice and claim within the provider's defined window, you may receive a credit against your next bill.

Every clause in that sentence is doing work.

The four parts

The target. A number, usually per month. Sometimes tiered by product or plan.

The definition. What counts as unavailable, measured how, from where, aggregated over what. This is the longest section and the one nobody reads.

The exclusions. What does not count. Also long.

The remedy. Almost always a service credit, on a sliding scale, capped.

Typical credit scale

Run the numbers on that. If you spend $1,000 a month and the provider delivers 99.5% - Three and a half hours of downtime - You are entitled to $100. If that outage cost your business $50,000 in lost transactions, you are still entitled to $100.

That is not a loophole. It is the explicit design: The credit is a proportional refund for service not delivered, not compensation for consequential loss. Every commercial SLA excludes consequential damages, and there is essentially no negotiating that at any normal contract size.

The exclusions that matter

Scheduled maintenance. Excluded universally. If a provider takes four hours of announced downtime, it does not count against the SLA at all.

Anything attributed to you. Your configuration, your code, your exceeded quotas, your failure to follow architecture guidance. This one is broader than it sounds: Cloud SLAs frequently require a multi-zone deployment to qualify, so a single-zone workload that goes down in a zone failure may have no claim.

Third parties and the internet. Network problems outside the provider's control, which includes a great deal of what users experience.

Beta and preview services. Usually no SLA at all.

Force majeure. Standard, and broadly drafted.

Credits are claimed, not granted

This is the most practically important thing on this page. Providers do not monitor your service level on your behalf and do not issue credits automatically.

The typical process: You notice the outage, you gather evidence with dates and times, you submit a claim through a specific channel, within a specific window - Often 30 days, sometimes as little as the end of the following billing cycle - And you calculate the credit yourself.

Miss the window and the entitlement is gone regardless of how bad the outage was. This is why most SLA credits are never claimed: The amounts are small, the process takes time, and the deadline passes.

If you intend to claim, keep your own monitoring records. A provider's status page is their account of events, and it is not always as granular as your experience. An independent record - A third-party outage log, your own monitoring, or both - Is considerably more persuasive.

SLA, SLO, SLI

Three terms that get used interchangeably and mean different things.

An SLI is a service level indicator: The actual measurement. Request success rate, latency at the 99th percentile.

An SLO is a service level objective: An internal target. "99.95% of requests succeed." Missing an SLO has engineering consequences - A team may freeze feature work to spend its error budget on reliability.

An SLA is the contract with the customer, and it is set lower than the internal SLO on purpose. A team targeting 99.95% internally might offer 99.9% externally, so that missing the internal target does not immediately create a contractual liability.

The gap between SLO and SLA is deliberate engineering headroom, and it is a sign of a mature provider rather than a cynical one.

What to actually do

Do not choose a provider on the SLA number. The remedy is too small to matter and the numbers are too similar to differentiate. Choose on the incident history, the quality and candour of the post-mortems, and how the status page behaves during a real event.

Read the exclusions before the target. They tell you more.

Architect as though there is no SLA, because functionally there is not. If downtime would genuinely damage you, the answer is redundancy across providers, not a contract clause.

Claim anyway. It takes twenty minutes and it is money you are owed. It also creates a record, which matters at renewal.