Self-hosted gateway maintenance

This guide covers node maintenance and connection drain settings for the enterprise gateway chart.

Keep replicas available

The enterprise chart creates a PodDisruptionBudget for the backplane. It permits one voluntary eviction at a time. Run at least two replicas to keep a ready replica during maintenance. A single replica can still be evicted.

Set the budget in your Helm values:

$terminalYAML

backplaneDisruptionBudget:
  enabled: true
  maxUnavailable: 1

The budget applies to node drains that use the Kubernetes eviction API. It does not stop direct pod deletion, process crashes, node failure, or a Deployment update. Configure the Deployment update strategy separately.

A drain can stop when no replacement pod can become Ready. Provide capacity on another node before you drain a node. Set backplaneDisruptionBudget.enabled to false only when your maintenance procedure permits concurrent evictions.

Set connection drain times

Set both the Proxy shutdown configuration and the chart's pod grace period:

$terminalYAML

apiVersion: core.apoxy.dev/v1alpha2
kind: Proxy
metadata:
  name: default
spec:
  provider: unmanaged
  shutdown:
    drainTimeout: 30s
    minimumDrainTime: 1s

Use the same drain timeout in your Helm values:

$terminalYAML

apoxy-gateway:
  backplane:
    proxy:
      drainTimeoutSeconds: 30

The chart adds five seconds to this value for terminationGracePeriodSeconds. The Proxy controls connection draining. The chart controls when Kubernetes forcibly stops the container. A Proxy edit does not update the pod grace period.

Verify

Read the budget and pod grace periods in your gateway namespace:

$terminalSH

kubectl --context <kube-context> -n <namespace> get pdb
kubectl --context <kube-context> -n <namespace> get pods \
  -l app.kubernetes.io/component=backplane \
  -o custom-columns=NAME:.metadata.name,GRACE:.spec.terminationGracePeriodSeconds

You should see one permitted disruption when all replicas are Ready, and a grace period of 35 seconds for the example configuration.