Run a container-backed service from a Cloudflare Worker with @cloudflare/containers, sizing instances and routing requests correctly

Documented steps

1. Confirm you are on the Workers Paid plan and that your image targets `linux/amd64`.
2. Install `@cloudflare/containers` and define the class: `export class MyContainer extends Container { defaultPort = 8080; sleepAfter = '10s'; envVars = { MESSAGE: '...' }; override onStart() {} override onStop() {} override onError(e: unknown) {} }`.
3. Add a `containers` entry in wrangler config specifying `image` (a Dockerfile path, a directory containing one, or a fully-qualified image reference), `class_name` (must match a Durable Object class), and `max_instances`.
4. Bind the backing Durable Object: `[[durable_objects.bindings]] name = "MY_CONTAINER" class_name = "MyContainer"`, plus a migration `[[migrations]] tag = "v1" new_sqlite_classes = ["MyContainer"]`. Containers require the SQLite-backed DO path, not `new_classes`.
5. For stateful or session-pinned work, route to a specific instance: `const c = env.MY_CONTAINER.getByName(sessionId); return await c.fetch(request);` — the same id always reaches the same container instance.
6. For interchangeable stateless work, load-balance instead: `const c = await getRandom(env.MY_CONTAINER, 3); return await c.fetch(request);`.
7. Ensure `defaultPort` matches the port your container process actually listens on — both fetch() and containerFetch() target it.
8. Pick an instance_type: lite (1/16 vCPU, 256 MiB, 2 GB disk), basic (1/4 vCPU, 1 GiB, 4 GB), standard-1 (1/2 vCPU, 4 GiB, 8 GB), standard-2 (1 vCPU, 6 GiB, 12 GB), standard-3 (2 vCPU, 8 GiB, 16 GB), standard-4 (4 vCPU, 12 GiB, 20 GB). `dev` and `standard` remain aliases for lite and standard-1.
9. For a custom size instead of a predefined type the constraints are: 1-4 vCPU, up to 12 GiB memory, up to 20 GB disk, minimum 3 GiB memory per vCPU, and at most 2 GB disk per 1 GiB memory.
10. Tune `sleepAfter` (e.g. '10s', '10m') to trade cold-start latency against paying for idle warm instances.
11. Handle shutdown: the platform sends SIGTERM and then SIGKILL after 15 minutes. Do graceful cleanup in `onStop()` / `onActivityExpired()`.
12. Deploy with `wrangler deploy`. Cloudflare uploads the image to its Registry and pre-fetches it globally; the first request starts an instance at the nearest location holding a pre-fetched image, and later requests to that instance route back to that same location while it stays alive.
13. Budget the billing model: charged per 10 ms of actively running instance time, memory and disk billed on provisioned resources, CPU billed on active use. Workers Paid includes 25 GiB-hours memory, 375 vCPU-minutes, and 200 GB-hours disk per month before overage — plus separate Workers, Durable Objects, and egress charges.
14. Official documentation (verify before relying on limits, which change): https://developers.cloudflare.com/containers/ | https://developers.cloudflare.com/containers/get-started/ | https://developers.cloudflare.com/containers/platform-details/ | https://developers.cloudflare.com/containers/platform-details/limits/ | https://developers.cloudflare.com/containers/pricing/