Four primitives, one API. Each one solves a specific way that shipping software to machines you don't control tends to break.
Publish once through the API and the build fans out to every edge node before your first client requests it. No manual push to individual regions, no waiting for a cache to warm on first request.
Binaries, container layers, and static payloads are kept byte-identical across every region. If a node falls behind during a deploy, it's excluded from routing until it catches up — never serves a half-synced artifact.
Clients on an older version pull only the binary diff, not the full package. On a typical point release this cuts payload size by 70–90%, which matters most on the connections that can least afford a full re-download.
Every node keeps the previous release on hand. A rollback call reverts a channel across the whole network in seconds, without waiting on a central coordinator to catch up first.
client device → nearest edge node (anycast routing, ~15–40ms) → local cache hit? yes → serve immediately no → pull from nearest synced peer → response streamed back, node caches for next request → origin is only touched on first publish, never per-request
The API reference has copy-paste examples for publishing your first release.
Read the docs