Walking through CloudFront’s edge caches, origin caches, TTL, and invalidation — plus region failover — is the operator’s mental model. A HIT on one edge is not freshness everywhere, and failover only helps if you know which mid-tier may fill after a miss. Self-hosted fleets hit the same shape once you add a shield hop or a second region: key inventory, who may fill origin, and coordinated purge. We have been building an open-source CDN control plane (CDNCA, AGPL-3.0) for self-hosted nodes and a plugin marketplace; marketplace features roll out release by release. In your deep dive, did TTL/invalidation tuning or region-failover behavior surprise people more once they traced a real miss?