Background

Design: WRITING-39975. Epic: SPM-4886.

Change

Build a Grafana dashboard to watch the safety of clusters with the feature on, and run a staged rollout: internal dev clusters, then production clusters in widening steps. Each step runs long enough to include step-ups. Move on only when the dashboard is clean. If it shows a problem, turn the kill switch off.

Metrics:

• layered_curs_blind_remove_follower rate per node: above zero on every follower where the feature is on.
• Orphan removes per cluster (layered_curs_blind_remove_ingest_deleted + disagg_step_up_ingest_drain_orphan_remove): near zero, alert when rising.
• layered_curs_blind_remove_leader: zero.
• Validation failures and index errors: zero, alert on any.
• Server asserts, crashes and restarts: no increase compared with clusters where the feature is off.

Acceptance criteria

• Dashboard and alerts are live. Each rollout step is recorded as clean before the next.

Depends on

MongoDB testing ticket (proctor override removal).