This document describes the changes needed to upgrade your Contour installation.
Contour 0.15 requires changes to your deployment manifests to explicitly opt in, or opt out of, secure communication between Contour and Envoy.
Contour 0.15 also adds experimental support for leader election which may be useful for installations which have split their Contour and Envoy containers into separate pods. A configuration we call split deployment.
Contour's contour serve
now requires that either TLS certificates be available, or you supply the --insecure
parameter.
If you do not supply TLS details or --insecure
, contour serve
will not start.
Due to the recently announced HTTP/2 vulnerabilities Contour 0.15 requires Envoy 1.11.1. As of August 2019, Envoy 1.11.1 is the only released version of Envoy that contains the fixes for those vulnerabilities.
Please see the Envoy Release Notes for information about issues fixed in Envoy 1.11.1.
If the following are true for you:
- Your installation is in the
heptio-contour
namespace. - You are using one of the example deployments.
- Your cluster can take few minutes of downtime.
Then the simplest way to upgrade to 0.15 is to delete the heptio-contour
namespace and reapply one of the example configurations.
From the root directory of the repository:
kubectl delete namespace heptio-contour
kubectl apply -f examples/<your-desired-deployment>
If you're using a LoadBalancer
Service, (which most of the examples do) deleting and recreating may change the public IP assigned by your cloud provider.
You'll need to re-check where your DNS names are pointing as well, using Get your hostname or IP address.
This section contains information for administrators who wish to apply the Contour 0.14 to 0.15 changes manually.
Change the Contour image version to gcr.io/heptio-images/contour:v0.15.0
.
You must either enable TLS for gRPC serving, or put --insecure
into your contour serve
startup line.
If you are running with both Contour and Envoy in a single pod, the existing deployment examples have already been updated with this change.
If you are running using the ds-hostnet-split
example or a derivative, we strongly recommend that you generate new certificates for securing your gRPC communication between Contour and Envoy.
There is a Job in the ds-hostnet-split
directory that will use the new contour certgen
command
to generate a CA and then sign Contour and Envoy keypairs, which can also then be saved directly
to Kubernetes as Secrets, ready to be mounted into your Contour and Envoy Deployments and Daemonsets.
If you would like more detail, see (grpc-tls-howto.md)[./grpc-tls-howto.md], which explains your options.
Contour 0.15 requires Envoy 1.11.1.
Change the Envoy image version to docker.io/envoyproxy/envoy:v1.11.1
.
Contour 0.15 adds experimental support for leader election.
Enabling leader election will mean that only one of the Contour pods will actually serve gRPC traffic.
This will ensure that all Envoy's take their configuration from the same Contour.
You can enable leader election with the --enable-leader-election
flag to contour serve
.
If you have deployed Contour and Envoy in their own pods--we call this split deployment--you should enable leader election so all envoy pods take their configuration from the lead contour.
To enable leader election, the following must be true for you:
- You are running in a split Contour and Envoy setup. That is, there are separate Contour and Envoy pod(s).
In order for leader election to work, you must make the following changes to your setup:
- The Contour Deployment must have its readiness probe changed too TCP readiness probe configured to check port 8001 (the gRPC port), as non-leaders will not serve gRPC, and Envoys may not be properly configured if they attempt to connect to a non-leader Contour. That is, you will need to change
readinessProbe:
httpGet:
path: /healthz
port: 8000
to
readinessProbe:
tcpSocket:
port: 8001
initialDelaySeconds: 15
periodSeconds: 10
inside the Pod spec.
- The update strategy for the Contour deployment must be changed to
Recreate
instead ofRollingUpdate
, as pods will never become Ready (since they won't pass the readiness probe). Add
strategy:
type: Recreate
to the top level of the Pod spec.
- Leader election is currently hard-coded to use a ConfigMap named
contour
in this namespace for the leader election lock. If you are using a newer installation of Contour, this may be present already, if not, the leader election library will create an empty ConfigMap for you.
Once these changes are made, add --enable-leader-election
to your contour serve
command. The
leader will perform and log its operations as normal, and the non-leaders will block waiting to
become leader. You can inspect the state of the leadership using
kubectl describe configmap -n heptio-contour contour
and checking the annotations that store exact details using
kubectl get configmap -n heptio-contour -o yaml contour