Reliability

Targets

• SLO: 99.9% availability per service. This is the internal target every service is built and measured against.
• SLA: 99% to customers. The gap between the 99.9% target and the 99% commitment is deliberate headroom, so a bad window does not immediately breach the customer contract.

How it is measured

Each service exports metrics to Prometheus, so availability and error rate are measured from real request outcomes rather than asserted. A typical success-rate query over the request and error counters:

sum(rate(requests_total{status!~"5.."}[30d]))
  / sum(rate(requests_total[30d]))

Fill in: the exact metric names your services expose, the scrape setup, and any alerting on the error budget.

Error budget

A 99.9% SLO allows roughly 43 minutes of downtime per 30-day window. Tracking burn against that budget is what turns the SLO from a number in a README into something operable: when the budget is healthy you ship, when it is nearly spent you slow down and stabilize.