Pod Disruption Budget Strategies

💡 Quick Answer: Use maxUnavailable: 1 for stateless services (allows one pod down during drain), minAvailable: "50%" for stateful workloads, and maxUnavailable: 0 for GPU training jobs that cannot tolerate any disruption.

The Problem

Node maintenance (upgrades, scaling, kernel patches) triggers pod evictions. Without PDBs, kubectl drain evicts all pods simultaneously — causing downtime. PDBs tell Kubernetes how many pods must remain available during voluntary disruptions.

The Solution

Strategy Matrix

Stateless Service

apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: web-pdb
spec:
  maxUnavailable: 1
  selector:
    matchLabels:
      app: web-frontend

Stateful Quorum (etcd, Kafka, ZooKeeper)

apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: etcd-pdb
spec:
  minAvailable: 2
  selector:
    matchLabels:
      app: etcd

For a 3-node etcd cluster: minAvailable: 2 ensures quorum (majority) is always maintained.

GPU Training — No Disruption

apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: training-pdb
spec:
  maxUnavailable: 0
  selector:
    matchLabels:
      app: distributed-training

⚠️ maxUnavailable: 0 blocks kubectl drain entirely until the training job completes. Use with unhealthyPodEvictionPolicy: AlwaysAllow (K8s 1.31+) to still evict crash-looping pods.

graph TD
    DRAIN[kubectl drain node] --> PDB{Check PDB}
    PDB -->|maxUnavailable: 1<br/>0 currently unavailable| EVICT[Evict 1 pod]
    PDB -->|maxUnavailable: 0| BLOCK[Block eviction<br/>Wait for completion]
    PDB -->|minAvailable: 2<br/>3 currently available| EVICT
    PDB -->|minAvailable: 2<br/>2 currently available| BLOCK
    EVICT --> RECHECK{Re-check PDB}
    RECHECK -->|Budget allows| EVICT
    RECHECK -->|Budget exhausted| WAIT[Wait for pods<br/>to become ready]

Common Issues

kubectl drain hangs forever

PDB with maxUnavailable: 0 or minAvailable equal to current replicas blocks all evictions. Use --timeout flag or temporarily adjust the PDB.

PDB blocks cluster autoscaler scale-down

Same issue — autoscaler respects PDBs. Ensure at least one pod can be disrupted for scale-down to work.

PDB selector doesn’t match any pods

Verify labels match: kubectl get pods -l app=web-frontend should return the pods you expect.

Best Practices

• Every production Deployment should have a PDB — no exceptions
• Use maxUnavailable for stateless — simpler, allows percentage-based scaling
• Use minAvailable for stateful — directly expresses quorum requirement
• Never set minAvailable = replicas — blocks all voluntary disruptions
• Combine with topologySpreadConstraints — spread pods across nodes so draining one node doesn’t violate PDB

Key Takeaways

• PDBs protect against voluntary disruptions only (node drain, cluster autoscaler) — not involuntary (OOM, hardware failure)
• maxUnavailable: 1 is the safest default for most services
• maxUnavailable: 0 blocks all evictions — use for jobs that cannot restart
• minAvailable with absolute numbers is safer than percentages for small replica counts
• PDBs don’t prevent scaling down replicas — only pod evictions