Fleet profiles

Both profiles run the compatible celld fork directly and use a dedicated S3 bucket. celld owns durability and recovery and tolerates the loss of any one member; the operator changes one member at a time. Bucket runs CELLD_DURABILITY=bucket: every acknowledged write is in S3 first, so its disk is a cache. PersistentFleet runs CELLD_DURABILITY=fleet: a write is acknowledged once its followers’ disks hold it.

Profile/layout	Workload and local disk	Planned contraction
Bucket / Deployment	Deployment with disk-backed emptyDir.	One member per step; Kubernetes chooses the deleted Pod.
Bucket / Ordered	StatefulSet with disk-backed emptyDir.	Highest ordinal, one per step.
PersistentFleet	StatefulSet with retained RWOP CSI disks.	Highest ordinal, one per step; its disk is kept and reattached if the fleet grows back.

Use Ordered Bucket when you need deterministic zone assignment by ordinal. Restart and upgrade roll one member at a time for both profiles. PersistentFleet keeps each member’s disk for the life of the fleet, including across scale-in. Its StatefulSet waits for each restarted member to be Ready before the next. It adds CSI requirements.

Profiles, Bucket layout, storage and placement are fixed at creation. There is no conversion between them, and no adoption of fleets or disks from the former operator or from another fleet. See storage configuration before choosing PersistentFleet.