What is request collapsing on a CDN?

Request collapsing is when a cache merges concurrent requests for the same object into a single request to the origin, then uses that one response for all of them. Fastly defines it this way. Without it, a popular page that misses in cache sends every concurrent visitor to the origin at once, the thundering herd problem.

100 requests for /pricing, cache empty
  without collapsing: up to 100 origin requests
  with collapsing:    1 origin request, 99 wait for it

Cloudflare calls its version a cache lock: one request at a time goes to the origin for a given asset from a single location. RFC 9111 allows the practice, and `Cache-Status` has a collapsed parameter to report it.

• It protects the origin, not the visitor. The waiting requests are still slow, because they wait for the first one.
• It works per location. Each PoP, or each upper tier with tiered caching, sends its own request.
• Responses that cannot be cached, such as private pages, cannot be shared, so collapsing does not help them.

After a purge, cache warming avoids the herd by fetching each page before the crowd arrives.

Documented request collapsing is one of the criteria in the Top 10 CDNs comparison; the Fastly, KeyCDN and Bunny.net guides describe three implementations.