Choosing Good SLIs
You'll learn what makes an SLI truly useful: being user-centric, measurable, and aligned with actual service health.
Now that you understand SLIs (Service Level Indicators) as the quantitative measurements of your service's behavior, the question becomes: which metrics should you choose? Not every metric deserves to be an SLI. The wrong choices lead to misleading dashboards, alert fatigue, and missed real problems.
Characteristics of Good SLIs
User-Centric
Your SLIs should reflect what users actually experience. Internal metrics like CPU usage or queue depth matter operationally, but users care about whether their requests succeed and how fast they return. A database might be at 90% CPU, but if requests still complete successfully in acceptable time, users are fine. Conversely, even with low CPU, a single failing dependency could break user journeys.
Think from the outside in: Would a user notice if this metric degraded?
Measurable
An SLI must be consistently, reliably, and affordably measured. "Page feels slow" isn't measurable; "95th percentile request latency over 500ms" is. You need concrete numbers you can track over time, ideally from production traffic itself.
Avoid metrics that are expensive to collect at scale or that introduce significant overhead — they'll be disabled when it matters most.
Aligned with Actual Service Health
Good SLIs predict problems users will report. If your SLI looks perfect but users are complaining, you chose the wrong indicator. Similarly, internal dashboards might show chaos that has zero user impact.
Test this alignment: when your SLI degrades, does service quality truly suffer? When users complain, does the corresponding SLI show it?
Key Takeaway: Effective SLIs are user-centric (reflect actual experience), measurable (concrete numbers you can track), and aligned with service health (degrade when users are actually affected). Choose metrics that answer "Are users happy?" not "Are servers busy?"