celld v0.4.0 now has a clear single-current deployment model: celld deploy moves deploy/current.json, then nodes adopt it in place.
What seems missing is the lifecycle of the immutable versions left behind. Previous deploy/<script>/<version> objects remain in the bucket, but there is no supported way to:
- list previous versions;
- access one directly without changing the current deployment;
- move current back to one for rollback;
- safely prune deployments and shared blobs no longer reachable through current or named pointers, queue attachments, or service bindings.
Direct version access might need an explicit contract for stateful applications: whether it runs against the fleet's live cell state or isolated preview state. Cloudflare's version preview URLs, for example, currently exclude Workers implementing Durable Objects. Still, someone picking a rollback version should be able to do so without having to guess whether the version is compatible with the current state of the fleet.
#175 asks for “a supported way to delete one cell”; this request concerns deployment artifacts instead.
Would a first-party version lifecycle like this be in scope? I have no strong preference on the exact CLI or endpoint shape, but I would prefer not to have external tooling guess celld's reachability and state rules
celld v0.4.0 now has a clear single-current deployment model:
celld deploymovesdeploy/current.json, then nodes adopt it in place.What seems missing is the lifecycle of the immutable versions left behind. Previous
deploy/<script>/<version>objects remain in the bucket, but there is no supported way to:Direct version access might need an explicit contract for stateful applications: whether it runs against the fleet's live cell state or isolated preview state. Cloudflare's version preview URLs, for example, currently exclude Workers implementing Durable Objects. Still, someone picking a rollback version should be able to do so without having to guess whether the version is compatible with the current state of the fleet.
#175 asks for “a supported way to delete one cell”; this request concerns deployment artifacts instead.
Would a first-party version lifecycle like this be in scope? I have no strong preference on the exact CLI or endpoint shape, but I would prefer not to have external tooling guess celld's reachability and state rules