In classic HTTP caching (RFC 7234), once an asset reaches its max-age TTL, it becomes completely expired. The next user request blocks synchronously while the edge proxy fetches the fresh resource from the origin backend (incurring 200ms to 500ms of latency). If thousands of users arrive simultaneously at expiration, this triggers a catastrophic "thundering herd" (cache stampede) that can crash origin databases. RFC 5861 introduces Cache-Control: max-age=600, stale-while-revalidate=86400. If a request arrives between 600 seconds and 87,000 seconds, the CDN edge serves the stale cached response IMMEDIATELY (in sub-15ms) to the user, and asynchronously triggers a single background fetch to origin. The user experiences zero latency penalty, origin load remains completely flat, and subsequent users receive the refreshed asset.

Legacy URL purges require enumerating every exact URL path and query string permutation (e.g. /product/iphone, /product/iphone?currency=EUR, /api/v2/items/912). If an e-commerce catalog contains 50,000 localized variants or dynamic pages featuring a single product, updating that product's price requires issuing 50,000 separate purge requests or wiping the entire site cache. Cache-Tags / Surrogate-Keys solve this by associating metadata tags with HTTP responses via edge headers (e.g. Surrogate-Key: prod_102 brand_apple cat_phones). When inventory changes, the origin issues a single API call: PURGE tag=prod_102. The CDN instantly marks every cached asset containing that tag as invalid across all global POPs in less than 150 milliseconds, regardless of URL variations.

A Hard Purge completely evicts the target object from edge storage and RAM immediately. The next incoming request MUST wait for origin revalidation, exposing origin servers to traffic spikes. A Soft Purge (stale purge) does not delete the asset: it immediately marks the asset as STALE. The edge CDN continues serving the stale asset to incoming users while asynchronously making a background conditional request (If-None-Match / If-Modified-Since) to origin. This guarantees uninterrupted high availability and sub-20ms user latency even during rapid price or catalog updates.

When an asset is completely missing from cache (cache miss) and 10,000 clients request it in the same second, naive proxies send 10,000 requests to the origin. Modern edge CDNs implement Request Collapsing (waiting lists): the edge node forwards the FIRST request to the origin and holds all subsequent 9,999 requests in local memory. When the origin returns the 200 OK, the edge node broadcasts the single response body to all 10,000 waiting clients simultaneously, collapsing 10,000 origin queries into exactly one.