Cluster fencing 101: why it prevents split-brain

Last updated: 20 Aug 2025

Definition: Fencing is how a cluster forcibly removes a failed/partitioned node to protect data. No fencing, no integrity.

Types

  • Power fencing (IPMI/iDRAC/iLO, PDUs): power-off the node.
  • Fabric fencing (SAN switch, SBD): cut disk or path access.

Minimal Pacemaker baseline

# Enable fencing
crm configure property stonith-enabled=true

# Example IPMI fence device (adjust params)
crm configure primitive fence-node1 stonith:external/ipmi \
  params ipaddr=10.0.10.11 login=admin passwd=REDACTED \
  op monitor interval=60s

Runbook (operators)

  1. Confirm node state; do not manually power-on a fenced node until cleared.
  2. Check logs: why fenced? power? storage? comms?
  3. Recover node offline; rejoin only after health checks pass.

Testing

  • Simulate node failure in maintenance window; verify STONITH executes.
  • Prove service failover; record timings.