What changed
On September 30, 2026 Cloudflare introduced the `durable_object` scheduling policy for Containers in public beta. Instead of a centrally configured image and instance type for an application, a Durable Object can select an immutable image reference and instance size when it starts a Container. Cloudflare also documented filesystem snapshots for Containers using this policy, allowing point-in-time filesystem state to be captured and restored later.
Why it matters
This gives builders a more stateful, individually addressable compute model inside Cloudflare's platform. Agent sandboxes, development environments and other long-lived workloads can choose runtime resources per instance and preserve disk state without coupling every instance to an application-wide image rollout. It narrows the gap between serverless orchestration and VM-like lifecycle control, while retaining important constraints around image compatibility and snapshot scope.
A Durable Object can now decide what Container to run
With the new scheduling policy, Wrangler prepares configured images and exposes immutable references through the Durable Object Container API. The Durable Object supplies the image and instance size when starting the Container. Cloudflare also provides a managed Debian Trixie image with Node.js 24.20.0 that can be started without defining a custom image.
Instances stop participating in one application-wide rollout
Cloudflare says Durable Object-managed Container instances have independent lifecycles and do not participate in application-wide image rollouts. That gives an application more control over when individual stateful workloads move between images or sizes, but it also means builders need an explicit update strategy for those instances.
Snapshots preserve disk state, not a running process
Snapshots capture the Container's complete filesystem state but not memory or running processes. They are immutable and tied to the Container image version that created them. A snapshot cannot be restored onto a different image version, and applications must use the durable_object scheduling policy to create or restore one.