Multi-cluster fleet

Kollect uses one ordinary operator per cluster. Operators write to a shared sink and partition rows by spec.cluster; there is no hub controller or ingest tier to operate.

flowchart LR
  A["Cluster A operator"] --> S["Shared sink"]
  B["Cluster B operator"] --> S
  C["Cluster C operator"] --> S

Use unique cluster identifiers and a layout that prevents writers from contending for the same path. For relational stores, include the cluster dimension in keys and indexes. Git fleets should partition paths and choose a cadence that avoids needless repository lock contention.

Snapshot layouts use {cluster} in their path templates. Database rows include cluster in their identity, and event consumers receive the same dimension. CR status keeps only summaries; there is no in-cluster aggregation tier or cross-cluster control plane.

See ADR-0501 and the multi-cluster example.