Redeploying Envoy

The Envoy process, the data path component of Contour, at times needs to be re-deployed. This could be due to an upgrade, a change in configuration, or a node-failure forcing a redeployment.

When implementing this roll out, the following steps should be taken:

1. Stop Envoy from accepting new connections
2. Start draining existing connections in Envoy by sending a POST request to /healthcheck/fail endpoint
3. Wait for connections to drain before allowing Kubernetes to SIGTERM the pod

Overview

Contour implements a new envoy sub-command which has a shutdown-manager whose job is to manage a single Envoy instances lifecycle for Kubernetes. The shutdown-manager runs as a new container alongside the Envoy container in the same pod. It exposes two HTTP endpoints which are used for livenessProbe as well as to handle the Kubernetes preStop event hook.

• livenessProbe: This is used to validate the shutdown manager is still running properly. If requests to /healthz fail, the container will be restarted.
• preStop: This is used to keep the container running while waiting for Envoy to drain connections. The /shutdown endpoint blocks until the connections are drained.

 - name: shutdown-manager
   command:
   - /bin/contour
   args:
     - envoy
     - shutdown-manager
   image: docker.io/projectcontour/contour:master
   imagePullPolicy: Always
   lifecycle:
     preStop:
       httpGet:
         path: /shutdown
         port: 8090
         scheme: HTTP
   livenessProbe:
     httpGet:
       path: /healthz
       port: 8090
     initialDelaySeconds: 3
     periodSeconds: 10  

The Envoy container also has some configuration to implement the shutdown manager. First the preStop hook is configured to use the /shutdown endpoint which blocks the container from exiting. Finally, the pod’s terminationGracePeriodSeconds is customized to extend the time in which Kubernetes will allow the pod to be in the Terminating state. The termination grace period defines an upper bound for long-lived sessions. If during shutdown, the connections aren’t drained to the configured amount, the terminationGracePeriodSeconds will send a SIGTERM to the pod killing it.

Shutdown Manager Config Options

The shutdown manager has a set of arguments that can be passed to change how it behaves:

• check-interval: Time to poll Envoy for open connections.
  • Type: duration (Default 5s)
• check-delay: Time wait before polling Envoy for open connections.
  • Type: duration (Default 60s)
• min-open-connections: Min number of open connections when polling Envoy.
  • Type: integer (Default 0)
• serve-port: Port to serve the http server on.
  • Type: integer (Default 8090)