Scale a fleet

Change the CelldFleet target, never its child Deployment or StatefulSet:

kubectl --context YOUR_CONTEXT -n fleets patch celldfleet my-fleet   --type merge -p '{"spec":{"replicas":4}}'
kubectl --context YOUR_CONTEXT -n fleets get celldfleet my-fleet -o yaml

Keep replicas at least equal to the configured zone count. Additions require eligible node capacity; PersistentFleet also needs new CSI claims for ordinals that have not had a member before.

Bucket fleets raise replicas directly. Scale-in lowers replicas by one member per step and waits for updated, ready replicas before the next. Ordered removes the highest ordinal; a Deployment lets Kubernetes choose the victim, and the member drains on SIGTERM. Automatic and External contraction also require survivor capacity for every possible victim, else CapacityUncertain.

PersistentFleet scale-in follows the same rule: one member per step, the highest ordinal, and only after the previous change has rolled out. The removed member drains on SIGTERM and its claim, such as data-my-fleet-3, is kept. An ordinal that had a member before reattaches its kept claim, which celld treats as a restart on the same disk; new ordinals get new claims. Growth adds one run of ordinals at a time, kept claims first when the fleet grows back, and each run waits until the scheduler has placed, or found no node for, every current member, so a new disk cannot take the zone a kept disk needs. A claim at a member’s name without this fleet’s UID label stops growth with StorageIdentityConflict. See the PersistentFleet lifecycle.

A removed member’s disk costs storage until the fleet grows back to that ordinal or is deleted. Delete its claim by hand only if the fleet will not grow back. celld stops depending on a departed member once its leaders re-form their ensembles, usually within seconds of the removal; a claim deleted before then can still hold the only copy of recent writes.

Read conditions while scaling runs. An external autoscaler must target the CelldFleet /scale subresource using capacity.mode: External. It uses the same checks and gains no direct workload authority. See capacity policy.