Authorization Policy Trust Domain Migration

4 minute read page test

This task shows you how to migrate from one trust domain to another without changing authorization policy.

In Istio 1.4, we introduce an alpha feature to support trust domain migration for authorization policy. This means if an Istio mesh needs to change its trust domain, the authorization policy doesn’t need to be changed manually. In Istio, if a workload is running in namespace foo with the service account bar, and the trust domain of the system is my-td, the identity of said workload is spiffe://my-td/ns/foo/sa/bar. By default, the Istio mesh trust domain is cluster.local, unless you specify it during the installation.

Before you begin

Read the authorization concept guide.

Install Istio with a custom trust domain and mutual TLS enabled.

$ istioctl install --set profile=demo --set values.global.trustDomain=old-td

Deploy the httpbin sample in the default namespace and the sleep sample in the default and sleep-allow namespaces:

$ kubectl label namespace default istio-injection=enabled
$ kubectl apply -f @samples/httpbin/httpbin.yaml@
$ kubectl apply -f @samples/sleep/sleep.yaml@
$ kubectl create namespace sleep-allow
$ kubectl label namespace sleep-allow istio-injection=enabled
$ kubectl apply -f @samples/sleep/sleep.yaml@ -n sleep-allow

Apply the authorization policy below to deny all requests to httpbin except from sleep in the sleep-allow namespace.

$ kubectl apply -f - <<EOF
apiVersion: security.istio.io/v1beta1
kind: AuthorizationPolicy
metadata:
  name: service-httpbin.default.svc.cluster.local
  namespace: default
spec:
  rules:
  - from:
    - source:
        principals:
        - old-td/ns/sleep-allow/sa/sleep
    to:
    - operation:
        methods:
        - GET
  selector:
    matchLabels:
      app: httpbin
---
EOF

Notice that it may take tens of seconds for the authorization policy to be propagated to the sidecars.

Verify that requests to httpbin from:

sleep in the default namespace are denied.

$ kubectl exec "$(kubectl get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -- curl http://httpbin.default:8000/ip -s -o /dev/null -w "%{http_code}\n"
403

sleep in the sleep-allow namespace are allowed.

$ kubectl exec "$(kubectl -n sleep-allow get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -n sleep-allow -- curl http://httpbin.default:8000/ip -s -o /dev/null -w "%{http_code}\n"
200

Migrate trust domain without trust domain aliases

Install Istio with a new trust domain.
```
$ istioctl install --set profile=demo --set values.global.trustDomain=new-td
```
Istio mesh is now running with a new trust domain, new-td.
Redeploy the httpbin and sleep applications to pick up changes from the new Istio control plane.
```
$ kubectl delete pod --all
```
```
$ kubectl delete pod --all -n sleep-allow
```
Verify that requests to httpbin from both sleep in default namespace and sleep-allow namespace are denied.
```
$ kubectl exec "$(kubectl get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -- curl http://httpbin.default:8000/ip -s -o /dev/null -w "%{http_code}\n"
403
```
```
$ kubectl exec "$(kubectl -n sleep-allow get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -n sleep-allow -- curl http://httpbin.default:8000/ip -s -o /dev/null -w "%{http_code}\n"
403
```
This is because we specified an authorization policy that deny all requests to httpbin, except the ones the old-td/ns/sleep-allow/sa/sleep identity, which is the old identity of the sleep application in sleep-allow namespace. When we migrated to a new trust domain above, i.e. new-td, the identity of this sleep application is now new-td/ns/sleep-allow/sa/sleep, which is not the same as old-td/ns/sleep-allow/sa/sleep. Therefore, requests from the sleep application in sleep-allow namespace to httpbin were allowed before are now being denied. Prior to Istio 1.4, the only way to make this work is to change the authorization policy manually. In Istio 1.4, we introduce an easy way, as shown below.

Migrate trust domain with trust domain aliases

Install Istio with a new trust domain and trust domain aliases.

$ cat <<EOF > ./td-installation.yaml
apiVersion: install.istio.io/v1alpha1
kind: IstioOperator
spec:
  values:
    global:
      trustDomain: new-td
      trustDomainAliases:
        - old-td
EOF
$ istioctl install --set profile=demo -f td-installation.yaml

Without changing the authorization policy, verify that requests to httpbin from:

sleep in the default namespace are denied.

$ kubectl exec "$(kubectl get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -- curl http://httpbin.default:8000/ip -s -o /dev/null -w "%{http_code}\n"
403

sleep in the sleep-allow namespace are allowed.

$ kubectl exec "$(kubectl -n sleep-allow get pod -l app=sleep -o jsonpath={.items..metadata.name})" -c sleep -n sleep-allow -- curl http://httpbin.default:8000/ip -s -o /dev/null -w "%{http_code}\n"
200

Best practices

Starting from Istio 1.4, when writing authorization policy, you should consider using the value cluster.local as the trust domain part in the policy. For example, instead of old-td/ns/sleep-allow/sa/sleep, it should be cluster.local/ns/sleep-allow/sa/sleep. Notice that in this case, cluster.local is not the Istio mesh trust domain (the trust domain is still old-td). However, in authorization policy, cluster.local is a pointer that points to the current trust domain, i.e. old-td (and later new-td), as well as its aliases. By using cluster.local in the authorization policy, when you migrate to a new trust domain, Istio will detect this and treat the new trust domain as the old trust domain without you having to include the aliases.

Clean up

$ kubectl delete authorizationpolicy service-httpbin.default.svc.cluster.local
$ kubectl delete deploy httpbin; kubectl delete service httpbin; kubectl delete serviceaccount httpbin
$ kubectl delete deploy sleep; kubectl delete service sleep; kubectl delete serviceaccount sleep
$ kubectl delete namespace sleep-allow
$ istioctl manifest generate --set profile=demo -f td-installation.yaml | kubectl delete --ignore-not-found=true -f -
$ rm ./td-installation.yaml