Salesforce's Argus team shares how they re-architected their internal observability platform to eliminate single-region AWS dependency while handling 4 billion metrics per minute. The original centralized model meant a single regional outage could take down global metrics visibility. Over six months, three teams redesigned the system into a geo-local model that processes and stores metrics closer to their origin, reducing cross-region data transfer costs and blast radius. A federation query layer backed by Elasticsearch routes requests only to relevant geographies, handles partial results gracefully, and returns HTTP 206 responses when a region is unavailable. The new architecture is live in 4 of 5 production regions, with OpenTSDB and HBase as core storage layers and metadata caching keeping wildcard query latency low.
