reliant-labs/forge/pdb

PodDisruptionBudgets

forge renders a PDB (maxUnavailable: 1) only above one replica. Two guards keep PDBs from blocking node drains:

• A PDB that can never allow a disruption is flagged — minAvailable >= replicas, maxUnavailable: 0, or minAvailable: 100%, on forge-rendered PDBs and your own forge.Manifests alike. forge lint reports it as pdb-blocks-drain (warning) and forge env deploy / up print a render-time warning naming the object. Fix: maxUnavailable: 1, or delete the PDB while the workload runs one replica.
• Stale PDBs are pruned. Dropping a workload to one replica stops rendering its PDB. Under Flux, prune: true removes it (PDBs are never prune-guarded). On the direct-apply path, forge env deploy deletes forge-labelled (managed-by=forge) PDBs of any workload in the render whose PDB the render no longer carries — PDBs only, and never one forge did not label. A PDB that belongs to no workload in the render at all (a replaced component's openbao-pdb, two budgets over the same pods) is caught by the owned-object prune instead: any PDB stamped forge.dev/owner that the render no longer carries goes once the rollout is healthy (deploy/plan).