Project Atlas

Globex's flagship API — 4 billion requests/month — was on a monolithic AWS setup that had grown organically over six years. Compute bills were $1.2M/month and rising. p95 latency at peak hit five seconds, and there had been three outages in the previous quarter.

We migrated the request path to Cloudflare Workers at the edge, with origin compute on AWS for stateful operations. Replaced six bespoke Kubernetes services with a uniform multi-region pattern. Introduced full OpenTelemetry observability, on-call runbooks, and a synthetic-monitoring suite.

The cost reductions came from three places: cold-start elimination at the edge, Postgres read replicas to offload heavy reads, and aggressive S3 lifecycle policies for the cold tail. We shipped region-by-region behind feature flags, with the old infrastructure kept warm during the cutover so rollback was one minute.