A content delivery system dynamically manages a content cache fleet by expanding or shrinking the size of the cache fleet to anticipate and/or respond to changes in demand for cached content. The content delivery system can consider various demand-based parameters when determining when and how to scale the cache fleet, including the overall demand (expected or observed) for all content available for delivery by the content delivery system, the demand for a subset of content or individual content items relative to the demand for other subsets of content or individual content items, etc. When content servers are removed from the cache fleet, snapshots of the content caches of the content servers can be stored to a persistent data store, and then restored to content servers when content servers are added to the cache fleet.