Choose a storage profile

Local storage requires the celld fork

If you use celld with persistent local storage (PersistentFleet), you must use the ewhauser/celld fork. It never counts an unreachable peer as holding no copy of a write, and it lets an idle leader stop depending on a departed member; stock upstream celld does neither. All runtime and recovery nodes must use a compatible fork. The fork is also required for Bucket fleets. See compatibility for the required release and image digest.

Read celld’s ownership and durability overview and durability guarantees to understand how it uses object storage and peer disks, when writes are acknowledged, and how recovery works. Those runtime tradeoffs should guide your choice of profile.

This page covers the Kubernetes configuration for that choice. Both profiles run celld directly and change one member at a time.

profile: Bucket runs CELLD_DURABILITY=bucket: every acknowledged write is in the bucket first, so the temporary local disk is a cache. Bucket does not need CSI or PVCs. Choose bucketWorkload: Ordered if you need deterministic zone assignment and highest-ordinal scale-in.

Choose profile: PersistentFleet when the runtime should use persistent local peer disks (CELLD_DURABILITY=fleet). Configure storage.storageClassName with a supported dynamic CSI class using Delete, WaitForFirstConsumer and RWOP claims. See storage for the full contract.

Setting	Bucket	PersistentFleet
Local disk	Disk-backed emptyDir, bounded by sizeGiB.	One RWOP claim per ordinal, kept for the life of the fleet.
Layout	Immutable Deployment or Ordered StatefulSet.	StatefulSet.
Scale-in	One member per step; highest ordinal for Ordered.	Highest ordinal, one per step; its disk is kept and reattached if the fleet grows back.
Restart/upgrade	Rolling, one member at a time.	Rolling, one member at a time on the same disk, highest ordinal first; each member must be Ready before the next.
PodDisruptionBudget	maxUnavailable: 1; node drains evict one member at a time.	maxUnavailable: 1; a member that is not Ready counts against it.

Neither profile provisions your bucket, runtime AWS identity, worker nodes or ingress controller. Optional routing can configure a public HTTPRoute or Ingress to an existing edge. Use one bucket per fleet. The reservation stays bound to the original fleet UID after deletion.

Profile and storage changes require a new fleet and bucket. There is no automatic conversion, disk adoption or compatibility path from the former operator.