Course contentsShow
← System Design
Lesson 1273 of 1,91932. Monitoring and AlertingFree lesson

Choosing Good SLIs

Learn to select SLIs that directly reflect user experience: availability, latency, throughput, and correctness.

Choosing Good SLIs

What you'll learn: How to select Service Level Indicators that meaningfully measure what your users actually experience.

Now that you understand what SLIs are, the critical question becomes: which indicators should you track? Not all metrics make good SLIs. The best SLIs directly reflect the quality of service from your users' perspective, not just system health.

The Four Pillars of User-Centric SLIs

Focus your SLI selection on these dimensions:

1. Availability

Can users access your service at all? This answers "Is it up?" Track the proportion of successful requests versus total requests.

2. Latency

How fast does your service respond? Measure request duration from the user's viewpoint. Focus on percentiles (like p99) rather than averages—one slow request matters to that one user.

3. Throughput

Can your system handle the user's workload? Track request rate and whether the system accepts or rejects traffic.

4. Correctness

Does your service return the right answer? This might mean tracking error rates, data validation failures, or consistency violations.

Selection Guidelines

Ask these questions:

  • Would users notice if this metric degraded?
  • Does it measure the outcome users care about, not just internal system state?
  • Can you act on it when it crosses a threshold?

Example: CPU utilization is a bad SLI—users don't care about your CPU. Request success rate is a good SLI—users definitely care if their requests fail.

Think of SLIs like a restaurant's quality measures: users care about wait time for food (latency), whether the kitchen is open (availability), if orders are correct (correctness), and if they can get a table (throughput)—not about the oven temperature.

Key Takeaway: Choose SLIs that directly measure user-facing outcomes—availability, latency, throughput, and correctness—not internal system metrics that users never experience.