SLI

Definition

An SLI, service level indicator, is the measurement behind an SLO: the specific, computable signal that stands in for whether users are having a good experience. The canonical form is a ratio, good events divided by total events, such as successful requests over all requests.

What it means in observability

The SLI is where reliability engineering succeeds or fails, because everything downstream, the objective, the error budget, the alerts, inherits its blind spots. A good SLI moves when users suffer and stays still when they do not. That sounds obvious and is genuinely difficult, since the metrics easiest to compute, CPU, uptime pings, server side status codes, are all imperfect proxies for what a person experienced at the other end of the request.

How it works in practice

SLIs are computed from telemetry. Availability SLIs count response outcomes, successes over total, with the definition of success chosen deliberately. Latency SLIs typically take a threshold form, the share of requests faster than a target, which behaves better in a ratio than a raw percentile does. Where you measure matters as much as what: load balancer measurements catch failures that in service measurements miss, since a crashed service cannot report its own errors.

Where it gets hard

Definition edge cases decide everything: do client errors count against you, do retries count once or twice, does a slow success count as good. Percentiles bring their own traps, since percentiles from different services or windows cannot simply be averaged. And missing data is the silent skew, because requests that never produced telemetry, often the worst failures, vanish from the denominator and flatter the number.

Where Tsuga fits

Tsuga computes SLI style queries in Analytics with ratio formulas and percentile aggregations across logs, metrics, and traces, and its flat per GB economics mean the underlying data is complete rather than a sample of itself.