Defining SLIs with custom metrics

Platform measurements such as load-balancer request counts may not capture the transactions an application needs to track. This episode shows how custom metrics become service-level indicators (SLIs), measurements of reliability, and how service-level objectives (SLOs) set targets for them 00:00–01:02(opens in a new tab).

Key Points Covered

• Prometheus reads an endpoint exposed by the application: The Node.js example counts events and records latency in a histogram. It exposes those values on a metrics endpoint that Prometheus scrapes—reads to collect measurements. A collector added to the Prometheus server forwards the data to Cloud Monitoring in the 2021 setup 01:02–03:08(opens in a new tab).
• A latency distribution can measure compliance with a threshold: A “distribution cut” counts the fraction of measured values within a selected range. The Prometheus example sets a target of 98% of responses completing within 500 milliseconds over a rolling 28-day period 02:04–03:08(opens in a new tab).
• Direct export is a different collection path: The separate OpenCensus example sends measurements from the application directly to Cloud Monitoring without scraping. OpenCensus is a legacy library; this is the video's historical alternative to maintaining a Prometheus server, not current library-selection advice 02:04–04:11(opens in a new tab).
• Choose histogram ranges around the question being asked: Buckets group measurements into ranges. The presenter recommends boundaries that distinguish requests inside and outside the SLO threshold, plus additional buckets for examining slow responses, such as the 99th-percentile latency 03:08–04:11(opens in a new tab).

Full video: https://www.youtube.com/watch?v=eCcsxDln6I0(opens in a new tab)