Design record: Architecture decisions

Design record

This page is for contributors and records a particular design or test snapshot. For current user guidance, see capabilities and limitations and the operation guides. Historical results do not establish current release or AWS validation.

0023: Node loss is routine. celld tolerates losing one node, so the operator proves no removals and never waits on a disk that is gone. Both profiles run celld directly with no launcher. Bucket fleets are plain rolling workloads, and voluntary disruption is one member at a time.

0024: PersistentFleet is a StatefulSet. The StatefulSet’s rolling update restarts members and celld recovers them. Every member keeps its disk until the fleet is deleted or the member cannot come back, the operator replaces such a member itself, and it keeps no lifecycle state.

Superseded records, including 0022’s strict removal proof, and their evidence are in Git history. The operator has no compatibility path for fleets created by those implementations beyond the in-place adoption described in 0024.