Cache-Control and Vary are the contract shared caches actually obey: what may be stored, how long it stays fresh, and which request dimensions fork the key. Get Vary wrong and you either fragment the cache into useless shards or serve the wrong variant across users. The ops half shows up when content changes — TTL alone is rarely enough; you still need a purge path that matches how the object was filled. Self-hosted edge fleets face the same inventory problem once more than one node holds a copy. We have been building CDNCA (AGPL-3.0) around CDN nodes and a plugin marketplace from one control plane; marketplace features roll out release by release. In the setups you see most often, does a careful Vary list or an explicit invalidation on publish prevent more stale-variant incidents?