What is an SLI?
An SLI (service level indicator) is a carefully defined measurement of how a service behaves for its users, usually good events divided by valid events, such as the share of HTTP requests that succeeded or finished within 300 ms. SLOs set targets on SLIs, so a badly chosen SLI makes every target built on it meaningless.
Why an SLI matters
"Is the service healthy?" needs a number everyone agrees on. An SLI fixes the definition: which requests count, what counts as good, and where it is measured (load balancer, application or synthetic check).
Good SLIs track what users feel: availability, latency, correctness and freshness. Server CPU is not an SLI; the checkout success rate is.
How to measure an SLI
SLI = good events ÷ valid events × 100% example definition: good = status < 500 and latency ≤ 300 ms; valid = all requests except health checks
Worked example
In one day an API received 1,000,000 valid requests (health checks excluded). 997,350 of them returned without a server error within 300 ms.
- SLI = 997,350 ÷ 1,000,000 = 0.99735, or 99.735%.
- Bad requests = 1,000,000 − 997,350 = 2,650.
- Against an SLO of 99.5% the day is fine (5,000 bad allowed); against 99.9% it is not, with 2,650 bad requests where 1,000 were allowed.
Result: The same measurement passes one target and fails another, which is why the SLI definition and the SLO must be agreed together.
Common mistakes
- Measuring what is easy. CPU and memory are easy to collect but are not what users experience.
- Counting health checks and bots. They inflate the denominator with trivially successful requests. Define which events are valid.
- Ratios without volume. 99% of 12 requests is not the same evidence as 99% of 12 million.
- Measuring only from inside. A server-side SLI misses DNS, TLS and CDN failures. Pair it with an external check.
How Vigil handles an SLI
- Uptime and API checks give an outside availability SLI per monitor (successful checks ÷ all checks), and the SLA report shows it against the target.
- APM endpoint data gives request-based indicators from inside: error rate, latency histograms and Apdex per endpoint.
- Triggers can alert on any of these metrics, with a warm-up and a cooldown.
SLI (service level indicator): frequently asked questions
What are the most common SLIs?
Availability (successful requests or checks), latency (the share of requests under a threshold), and for data systems, freshness and correctness.
Should an SLI be a percentage?
Expressing it as good ÷ valid events makes it easy to set targets and compute error budgets, so most teams do.
Where should SLIs be measured?
As close to the user as practical: at the load balancer or edge for request SLIs, plus external synthetic checks for full-path availability.