Fencing a Postgres primary that might still be alive

Managed Postgres services must fence old primary servers during failovers to prevent split-brain scenarios and lost writes. Because standard methods like stopping the cluster or updating HBA rules fail for existing connections or network partitions, Ubicloud's control plane uses environment-specific mechanisms including host veth rules, detached AWS user NICs, and deleted GCP Dynamic NICs. When all fencing approaches fail, the system intentionally proceeds with promotion to prioritize availability. Open article →