How does Consul Service Mesh Work on Kubernetes?

Consul service mesh automates service-to-service authorization and encryption across your Consul services. You can use service mesh in Kubernetes-orchestrated networks to secure communication between pods as well as communication between pods and external Kubernetes services.

Workflow

Consul service mesh is enabled by default when you install Consul on Kubernetes using the Consul Helm chart. Consul also automatically injects sidecars into the pods in your clusters that run Envoy. These sidecar proxies, called Consul dataplanes, are enabled when connectInject.default is set to false in the Helm chart. Refer to the following documentation for additional information about these concepts:

If connectInject.default is set to false or you want to explicitly enable service mesh sidecar proxy injection for a specific deployment, add the consul.hashicorp.com/connect-inject annotation to the pod specification template and set it to true when connecting services to the mesh.

Service names

When the service is onboarded, the name registered in Consul is set to the name of the Kubernetes Service associated with the Pod. You can use the consul.hashicorp.com/connect-service annotation to specify a custom name for the service, but if ACLs are enabled then the name of the service registered in Consul must match the Pod’s ServiceAccount name.

Transparent proxy mode

By default, the Consul service mesh runs in transparent proxy mode. This mode forces inbound and outbound traffic through the sidecar proxy even though the service binds to all interfaces. Transparent proxy infers the location of upstream services using Consul service intentions, and also allows you to use Kubernetes DNS as you normally would for your workloads.

When transparent proxy mode is enabled, all service-to-service traffic is required to use mTLS. When onboarding new services to service mesh, your network may have mixed mTLS and non-mTLS traffic, which can result in broken service-to-service communication. You can temporarily enable permissive mTLS mode during the onboarding process so that existing mesh services can accept traffic from services that are not yet fully onboarded. Permissive mTLS enables sidecar proxies to access both mTLS and non-mTLS traffic. Refer to Onboard mesh services in transparent proxy mode for additional information.

Kubernetes service mesh workload scenarios

Note: A Kubernetes Service is required in order to register services on the Consul service mesh. Consul monitors the lifecyle of the Kubernetes Service and its service instances using the service object. In addition, the Kubernetes service is used to register and de-register the service from Consul’s catalog.

The following configurations are examples for registering workloads on Kubernetes into Consul’s service mesh in different scenarios. Each scenario provides an example Kubernetes manifest to demonstrate how to use Consul’s service mesh with a specific Kubernetes workload type.

Kubernetes Pods running as a deployment

The following example shows a Kubernetes configuration that specifically enables service mesh connections for the static-server service. Consul starts and registers a sidecar proxy that listens on port 20000 by default and proxies valid inbound connections to port 8080.

Overview - 图1

static-server.yaml

  1. apiVersion: v1
  2. kind: Service
  3. metadata:
  4.   # This name will be the service name in Consul.
  5.   name: static-server
  6. spec:
  7.   selector:
  8.     app: static-server
  9.   ports:
  10.     - protocol: TCP
  11.       port: 80
  12.       targetPort: 8080
  13. ---
  14. apiVersion: v1
  15. kind: ServiceAccount
  16. metadata:
  17.   name: static-server
  18. ---
  19. apiVersion: apps/v1
  20. kind: Deployment
  21. metadata:
  22.   name: static-server
  23. spec:
  24.   replicas: 1
  25.   selector:
  26.     matchLabels:
  27.       app: static-server
  28.   template:
  29.     metadata:
  30.       name: static-server
  31.       labels:
  32.         app: static-server
  33.       annotations:
  34.         'consul.hashicorp.com/connect-inject': 'true'
  35.     spec:
  36.       containers:
  37.         - name: static-server
  38.           image: hashicorp/http-echo:latest
  39.           args:
  40.             - -text="hello world"
  41.             - -listen=:8080
  42.           ports:
  43.             - containerPort: 8080
  44.               name: http
  45.       # If ACLs are enabled, the serviceAccountName must match the Consul service name.
  46.       serviceAccountName: static-server

To establish a connection to the upstream Pod using service mesh, a client must dial the upstream workload using a mesh proxy. The client mesh proxy will use Consul service discovery to find all available upstream proxies and their public ports.

Connecting to mesh-enabled Services

The example Deployment specification below configures a Deployment that is capable of establishing connections to our previous example “static-server” service. The connection to this static text service happens over an authorized and encrypted connection via service mesh.

Overview - 图2

static-client.yaml

  1. apiVersion: v1
  2. kind: Service
  3. metadata:
  4.   # This name will be the service name in Consul.
  5.   name: static-client
  6. spec:
  7.   selector:
  8.     app: static-client
  9.   ports:
  10.     - port: 80
  11. ---
  12. apiVersion: v1
  13. kind: ServiceAccount
  14. metadata:
  15.   name: static-client
  16. ---
  17. apiVersion: apps/v1
  18. kind: Deployment
  19. metadata:
  20.   name: static-client
  21. spec:
  22.   replicas: 1
  23.   selector:
  24.     matchLabels:
  25.       app: static-client
  26.   template:
  27.     metadata:
  28.       name: static-client
  29.       labels:
  30.         app: static-client
  31.       annotations:
  32.         'consul.hashicorp.com/connect-inject': 'true'
  33.     spec:
  34.       containers:
  35.         - name: static-client
  36.           image: curlimages/curl:latest
  37.           # Just spin & wait forever, we'll use `kubectl exec` to demo
  38.           command: ['/bin/sh', '-c', '--']
  39.           args: ['while true; do sleep 30; done;']
  40.       # If ACLs are enabled, the serviceAccountName must match the Consul service name.
  41.       serviceAccountName: static-client

By default when ACLs are enabled or when ACLs default policy is allow, Consul will automatically configure proxies with all upstreams from the same datacenter. When ACLs are enabled with default deny policy, you must supply an intention to tell Consul which upstream you need to talk to.

When upstreams are specified explicitly with the consul.hashicorp.com/connect-service-upstreams annotation, the injector will also set environment variables <NAME>_CONNECT_SERVICE_HOST and <NAME>_CONNECT_SERVICE_PORT in every container in the Pod for every defined upstream. This is analogous to the standard Kubernetes service environment variables, but point instead to the correct local proxy port to establish connections via service mesh.

You can verify access to the static text server using kubectl exec. Because transparent proxy is enabled by default, use Kubernetes DNS to connect to your desired upstream.

  1. $ kubectl exec deploy/static-client -- curl --silent http://static-server/
  2. "hello world"

You can control access to the server using intentions. If you use the Consul UI or CLI to deny communication between “static-client” and “static-server”, connections are immediately rejected without updating either of the running pods. You can then remove this intention to allow connections again.

  1. $ kubectl exec deploy/static-client -- curl --silent http://static-server/
  2. command terminated with exit code 52

Kubernetes Jobs

Kubernetes Jobs run pods that only make outbound requests to services on the mesh and successfully terminate when they are complete. In order to register a Kubernetes Job with the mesh, you must provide an integer value for the consul.hashicorp.com/sidecar-proxy-lifecycle-shutdown-grace-period-seconds annotation. Then, issue a request to the http://127.0.0.1:20600/graceful_shutdown API endpoint so that Kubernetes gracefully shuts down the consul-dataplane sidecar after the job is complete.

Below is an example Kubernetes manifest that deploys a job correctly.

Overview - 图3

test-job.yaml

  1. ---
  2. apiVersion: v1
  3. kind: ServiceAccount
  4. metadata:
  5.   name: test-job
  6.   namespace: default
  7. ---
  8. apiVersion: v1
  9. kind: Service
  10. metadata:
  11.   name: test-job
  12.   namespace: default
  13. spec:
  14.   selector:
  15.     app: test-job
  16.   ports:
  17.     - port: 80
  18. ---
  19. apiVersion: batch/v1
  20. kind: Job
  21. metadata:
  22.    name: test-job
  23.    namespace: default
  24.    labels:
  25.      app: test-job
  26. spec:
  27.   template:
  28.     metadata:
  29.       annotations:
  30.         'consul.hashicorp.com/connect-inject': 'true'
  31.         'consul.hashicorp.com/sidecar-proxy-lifecycle-shutdown-grace-period-seconds': '5'
  32.       labels:
  33.         app: test-job
  34.     spec:
  35.       containers:
  36.       - name: test-job
  37.         image: alpine/curl:3.14
  38.         ports:
  39.         - containerPort: 80
  40.         command:
  41.           - /bin/sh
  42.           - -c
  43.           - |
  44.             echo "Started test job"
  45.             sleep 10
  46.             echo "Killing proxy"
  47.             curl --max-time 2 -s -f -XPOST http://127.0.0.1:20600/graceful_shutdown
  48.             sleep 10
  49.             echo "Ended test job"
  50.       serviceAccountName: test-job
  51.       restartPolicy: Never

Upon completing the job you should be able to verify that all containers are shut down within the pod.

  1. $ kubectl get pods  
  2. NAME             READY   STATUS      RESTARTS   AGE
  3. test-job-49st7   0/2     Completed   0          3m55s
  1. $ kubectl get job
  2. NAME       COMPLETIONS   DURATION   AGE
  3. test-job   1/1           30s        4m31s

In addition, based on the logs emitted by the pod you can verify that the proxy was shut down before the Job completed.

  1. $ kubectl logs test-job-49st7 -c test-job
  2. Started test job
  3. Killing proxy
  4. Ended test job

Kubernetes Pods with multiple ports

To configure a pod with multiple ports to be a part of the service mesh and receive and send service mesh traffic, you will need to add configuration so that a Consul service can be registered per port. This is because services in Consul currently support a single port per service instance.

In the following example, suppose we have a pod which exposes 2 ports, 8080 and 9090, both of which will need to receive service mesh traffic.

First, decide on the names for the two Consul services that will correspond to those ports. In this example, the user chooses the names web for 8080 and web-admin for 9090.

Create two service accounts for web and web-admin:

Overview - 图4

multiport-web-sa.yaml

  1. apiVersion: v1
  2. kind: ServiceAccount
  3. metadata:
  4.   name: web
  5. ---
  6. apiVersion: v1
  7. kind: ServiceAccount
  8. metadata:
  9.   name: web-admin

Create two Service objects for web and web-admin:

Overview - 图5

multiport-web-svc.yaml

  1. apiVersion: v1
  2. kind: Service
  3. metadata:
  4.   name: web
  5. spec:
  6.   selector:
  7.     app: web
  8.   ports:
  9.     - protocol: TCP
  10.       port: 80
  11.       targetPort: 8080
  12. ---
  13. apiVersion: v1
  14. kind: Service
  15. metadata:
  16.   name: web-admin
  17. spec:
  18.   selector:
  19.     app: web
  20.   ports:
  21.     - protocol: TCP
  22.       port: 80
  23.       targetPort: 9090

web will target containerPort 8080 and select pods labeled app: web. web-admin will target containerPort 9090 and will also select the same pods.

Kubernetes 1.24+ only In Kubernetes 1.24+ you need to create a Kubernetes secret for each multi-port service that references the ServiceAccount, and the Kubernetes secret must have the same name as the ServiceAccount:

Overview - 图6

multiport-web-secret.yaml

  1. apiVersion: v1
  2. kind: Secret
  3. metadata:
  4.   name: web
  5.   annotations:
  6.     kubernetes.io/service-account.name: web
  7.   type: kubernetes.io/service-account-token
  8. ---
  9. apiVersion: v1
  10. kind: Secret
  11. metadata:
  12.   name: web-admin
  13.   annotations:
  14.     kubernetes.io/service-account.name: web-admin
  15.   type: kubernetes.io/service-account-token

Create a Deployment with any chosen name, and use the following annotations:

  1. annotations:
  2.   'consul.hashicorp.com/connect-inject': 'true'
  3.   'consul.hashicorp.com/transparent-proxy': 'false'
  4.   'consul.hashicorp.com/connect-service': 'web,web-admin'
  5.   'consul.hashicorp.com/connect-service-port': '8080,9090'

Note that the order the ports are listed in the same order as the service names, i.e. the first service name web corresponds to the first port, 8080, and the second service name web-admin corresponds to the second port, 9090.

The service account on the pod spec for the deployment should be set to the first service name web:

  1. serviceAccountName: web

The following deployment example demonstrates the required annotations for the manifest. In addition, the previous YAML manifests can also be combined into a single manifest for easier deployment.

Overview - 图7

multiport-web.yaml

  1. apiVersion: apps/v1
  2. kind: Deployment
  3. metadata:
  4.   name: web
  5. spec:
  6.   replicas: 1
  7.   selector:
  8.     matchLabels:
  9.       app: web
  10.   template:
  11.     metadata:
  12.       name: web
  13.       labels:
  14.         app: web
  15.       annotations:
  16.         'consul.hashicorp.com/connect-inject': 'true'
  17.         'consul.hashicorp.com/transparent-proxy': 'false'
  18.         'consul.hashicorp.com/connect-service': 'web,web-admin'
  19.         'consul.hashicorp.com/connect-service-port': '8080,9090'
  20.     spec:
  21.       containers:
  22.         - name: web
  23.           image: hashicorp/http-echo:latest
  24.           args:
  25.             - -text="hello world"
  26.             - -listen=:8080
  27.           ports:
  28.             - containerPort: 8080
  29.               name: http
  30.         - name: web-admin
  31.           image: hashicorp/http-echo:latest
  32.           args:
  33.             - -text="hello world from 9090"
  34.             - -listen=:9090
  35.           ports:
  36.             - containerPort: 9090
  37.               name: http
  38.       serviceAccountName: web

After deploying the web application, you can test service mesh connections by deploying the static-client application with the configuration in the previous section and add the consul.hashicorp.com/connect-service-upstreams: 'web:1234,web-admin:2234' annotation to the pod template on static-client:

Overview - 图8

multiport-static-client.yaml

  1. 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142apiVersion: v1
  2. kind: Service
  3. metadata:
  4.   # This name will be the service name in Consul.
  5.   name: static-client
  6. spec:
  7.   selector:
  8.     app: static-client
  9.   ports:
  10.     - port: 80
  11. ---
  12. apiVersion: v1
  13. kind: ServiceAccount
  14. metadata:
  15.   name: static-client
  16. ---
  17. apiVersion: apps/v1
  18. kind: Deployment
  19. metadata:
  20.   name: static-client
  21. spec:
  22.   replicas: 1
  23.   selector:
  24.     matchLabels:
  25.       app: static-client
  26.   template:
  27.     metadata:
  28.       name: static-client
  29.       labels:
  30.         app: static-client
  31.       annotations:
  32.         'consul.hashicorp.com/connect-inject': 'true'
  33.         'consul.hashicorp.com/connect-service-upstreams': 'web:1234,web-admin:2234'
  34.     spec:
  35.       containers:
  36.         - name: static-client
  37.           image: curlimages/curl:latest
  38.           # Just spin & wait forever, we'll use `kubectl exec` to demo
  39.           command: ['/bin/sh', '-c', '--']
  40.           args: ['while true; do sleep 30; done;']
  41.       # If ACLs are enabled, the serviceAccountName must match the Consul service name.
  42.       serviceAccountName: static-client

If you exec on to a static-client pod, using a command like:

  1. $ kubectl exec -it static-client-5bd667fbd6-kk6xs -- /bin/sh

you can then run:

  1. $ curl localhost:1234

to see the output hello world and run:

  1. $ curl localhost:2234

to see the output hello world from 9090.

The way this works is that a Consul service instance is being registered per port on the Pod, so there are 2 Consul services in this case. An additional Envoy sidecar proxy and connect-init init container are also deployed per port in the Pod. So the upstream configuration can use the individual service names to reach each port as seen in the example.

Caveats for Multi-port Pods

  • Transparent proxy is not supported for multi-port Pods.
  • Metrics and metrics merging is not supported for multi-port Pods.
  • Upstreams will only be set on the first service’s Envoy sidecar proxy for the pod.
    • This means that ServiceIntentions from a multi-port pod to elsewhere, will need to use the first service’s name, web in the example above to accept connections from either web or web-admin. ServiceIntentions from elsewhere to a multi-port pod can use the individual service names within the multi-port Pod.
  • Health checking is done on a per-Pod basis, so if any Kubernetes health checks (like readiness, liveness, etc) are failing for any container on the Pod, the entire Pod is marked unhealthy, and any Consul service referencing that Pod will also be marked as unhealthy. So, if web has a failing health check, web-admin would also be marked as unhealthy for service mesh traffic.

Installation and Configuration

The service mesh sidecar proxy is injected via a mutating admission webhook call the connect injector provided by the consul-k8s project. This enables the automatic pod mutation shown in the usage section above. Installation of the mutating admission webhook is automated using the Helm chart.

To install the connect injector, enable the connect injection feature using Helm values and upgrade the installation using helm upgrade for existing installs or helm install for a fresh install.

  1. connectInject:
  2.   enabled: true

This will configure the injector to inject when the injection annotation is set to true. Other values in the Helm chart can be used to limit the namespaces the injector runs in, enable injection by default, and more.

Verifying the Installation

To verify the installation, run the “Accepting Inbound Connections” example from the “Usage” section above. After running this example, run kubectl get pod static-server --output yaml. In the raw YAML output, you should see connect injected containers and an annotation consul.hashicorp.com/connect-inject-status set to injected. This confirms that injection is working properly.

If you do not see this, then use kubectl logs against the injector pod and note any errors.

Controlling Injection Via Annotation

By default, the injector will inject only when the injection annotation on the pod (not the deployment) is set to true:

  1. annotations:
  2.   'consul.hashicorp.com/connect-inject': 'true'

Injection Defaults

If you wish for the injector to always inject, you can set the default to true in the Helm chart:

  1. connectInject:
  2.   enabled: true
  3.   default: true

You can then exclude specific pods via annotation:

  1. annotations:
  2.   'consul.hashicorp.com/connect-inject': 'false'

Controlling Injection Via Namespace

You can control which Kubernetes namespaces are allowed to be injected via the k8sAllowNamespaces and k8sDenyNamespaces keys:

  1. connectInject:
  2.   enabled: true
  3.   k8sAllowNamespaces: ['*']
  4.   k8sDenyNamespaces: []

In the default configuration (shown above), services from all namespaces are allowed to be injected. Whether or not they’re injected depends on the value of connectInject.default and the consul.hashicorp.com/connect-inject annotation.

If you wish to only enable injection in specific namespaces, you can list only those namespaces in the k8sAllowNamespaces key. In the configuration below only the my-ns-1 and my-ns-2 namespaces will be enabled for injection. All other namespaces will be ignored, even if the connect inject annotation is set.

  1. connectInject:
  2.   enabled: true
  3.   k8sAllowNamespaces: ['my-ns-1', 'my-ns-2']
  4.   k8sDenyNamespaces: []

If you wish to enable injection in every namespace except specific namespaces, you can use * in the allow list to allow all namespaces and then specify the namespaces to exclude in the deny list:

  1. connectInject:
  2.   enabled: true
  3.   k8sAllowNamespaces: ['*']
  4.   k8sDenyNamespaces: ['no-inject-ns-1', 'no-inject-ns-2']

NOTE: The deny list takes precedence over the allow list. If a namespace is listed in both lists, it will not be synced.

NOTE: The kube-system and kube-public namespaces will never be injected.

Consul Enterprise Namespaces

Consul Enterprise 1.7+ supports Consul namespaces. When Kubernetes pods are registered into Consul, you can control which Consul namespace they are registered into.

There are three options available:

  1. Single Destination Namespace – Register all Kubernetes pods, regardless of namespace, into the same Consul namespace.

    This can be configured with:

    1. global:
    2.   enableConsulNamespaces: true
    4. connectInject:
    5.   enabled: true
    6.   consulNamespaces:
    7.     consulDestinationNamespace: 'my-consul-ns'

    NOTE: If the destination namespace does not exist we will create it.

  2. Mirror Namespaces - Register each Kubernetes pod into a Consul namespace with the same name as its Kubernetes namespace. For example, pod foo in Kubernetes namespace ns-1 will be synced to the Consul namespace ns-1. If a mirrored namespace does not exist in Consul, it will be created.

    This can be configured with:

    1. global:
    2.   enableConsulNamespaces: true
    4. connectInject:
    5.   enabled: true
    6.   consulNamespaces:
    7.     mirroringK8S: true

  3. Mirror Namespaces With Prefix - Register each Kubernetes pod into a Consul namespace with the same name as its Kubernetes namespace with a prefix. For example, given a prefix k8s-, pod foo in Kubernetes namespace ns-1 will be synced to the Consul namespace k8s-ns-1.

    This can be configured with:

    1. global:
    2.   enableConsulNamespaces: true
    4. connectInject:
    5.   enabled: true
    6.   consulNamespaces:
    7.     mirroringK8S: true
    8.     mirroringK8SPrefix: 'k8s-'

Consul Enterprise Namespace Upstreams

When transparent proxy is enabled and ACLs are disabled, the upstreams will be configured automatically across Consul namespaces. When ACLs are enabled, you must configure it by specifying an intention, allowing services across Consul namespaces to talk to each other.

If you wish to specify an upstream explicitly via the consul.hashicorp.com/connect-service-upstreams annotation, use the format [service-name].[namespace]:[port]:[optional datacenter]:

  1. annotations:
  2.   'consul.hashicorp.com/connect-inject': 'true'
  3.   'consul.hashicorp.com/connect-service-upstreams': '[service-name].[namespace]:[port]:[optional datacenter]'

See consul.hashicorp.com/connect-service-upstreams for more details.

Note: When you specify upstreams via an upstreams annotation, you will need to use localhost:<port> with the port from the upstreams annotation instead of KubeDNS to connect to your upstream application.