Engineering for Reliability
Defining SLIs with platform metrics
Twenty-three substantive episodes on SLOs, telemetry, alerting, investigation, and operations
Twenty-three 2021–2022 episodes connect user-facing reliability targets to metrics, logs, traces, and alerts. Examples show instrumentation, Prometheus collection, quota monitoring, GKE investigation, and observability cost controls.
In this collection Browse 23 summaries 19 of 23
Existing platform metrics can measure reliability without adding instrumentation to application code. This episode builds service-level indicators (SLIs), measurements of user-visible success, from load-balancer traffic, data-pipeline lag, and request logs, then sets service-level objectives (SLOs) as targets for those measurements 00:00–01:02(opens in a new tab), 07:19–09:25(opens in a new tab).
Key Points Covered
- Count successful requests or successful time intervals: A request-based SLI divides good units of work by total work. A window-based SLI instead measures the fraction of time intervals in which the service meets a threshold. The available measurement determines which kind of ratio can be calculated 00:00–01:02(opens in a new tab), 05:13–06:17(opens in a new tab).
- Filters encode product judgment: The load-balancer example counts HTTP 200 responses as good and asks teams to decide deliberately whether responses such as 404s belong in the total 02:06–04:11(opens in a new tab).
- Use a windowed target when individual events cannot be counted: The Dataflow example lacks per-element good-versus-total counts. Instead, it requires 99% of five-minute windows over 28 days to have average system lag below two seconds. This measures the fraction of acceptable windows, not the fraction of individual messages processed in under two seconds 05:13–07:19(opens in a new tab).
- Turn request logs into counts: When the application logs a status for each request, a logs-based metric can count total requests and filter successful ones by status. The demonstrated availability SLI divides HTTP 200 responses by all requests 07:19–09:25(opens in a new tab).
Full video: https://www.youtube.com/watch?v=uhpAScSerec(opens in a new tab)