Observability at Scale for a Global Retailer

Prometheus solves metrics collection well and long-term storage badly. Its local time-series database is built for weeks of retention on a single node, which left the retailer facing three problems simultaneously. Retention: seasonal retail analysis depends on year-over-year comparison, and a fortnight of history cannot answer whether this peak trading period behaved like the last one. Availability: a single Prometheus is a single point of failure, yet running redundant replicas produces two divergent datasets rather than one authoritative view. Locality: with infrastructure spread across regions, each Prometheus sees only its own slice, so any platform-wide question means querying every instance and reconciling the answers by hand. The retailer needed one query surface across all of it, plus dashboards and alerting that on-call engineers would actually trust at three in the morning.