Transparent Proxy

Transparent proxy allows applications to communicate through the mesh without changing their configuration. Transparent proxy also hardens application security by preventing direct inbound connections that bypass the mesh.

Without Transparent Proxy

Diagram demonstrating that without transparent proxy, applications must "opt in" to connecting to their dependencies through the mesh

Without transparent proxy, application owners need to:

  1. Explicitly configure upstream services, choosing a local port to access them.
  2. Change application to access localhost:<chosen port>.
  3. Configure application to listen only on the loopback interface to prevent unauthorized traffic from bypassing the mesh.

With Transparent Proxy

Diagram demonstrating that with transparent proxy, connections are automatically routed through the mesh

With transparent proxy:

  1. Upstreams are inferred from service intentions, so no explicit configuration is needed.
  2. Outbound connections pointing to a KubeDNS name “just work” — network rules redirect them through the proxy.
  3. Inbound traffic is forced to go through the proxy to prevent unauthorized direct access to the application.

Overview

Transparent proxy allows users to reach other services in the service mesh while ensuring that inbound and outbound traffic for services in the mesh are directed through the sidecar proxy. Traffic is secured and only reaches intended destinations since the proxy can enforce security and policy like TLS and Service Intentions.

Previously, service mesh users would need to explicitly define upstreams for a service as a local listener on the sidecar proxy, and dial the local listener to reach the appropriate upstream. Users would also have to set intentions to allow specific services to talk to one another. Transparent proxying reduces this duplication, by determining upstreams implicitly from Service Intentions. Explicit upstreams are still supported in the proxy service registration on VMs and via the annotation in Kubernetes.

To support transparent proxying, Consul’s CLI now has a command consul connect redirect-traffic to redirect traffic through an inbound and outbound listener on the sidecar. Consul also watches Service Intentions and configures the Envoy proxy with the appropriate upstream IPs. If the default ACL policy is “allow”, then Service Intentions are not required. In Consul on Kubernetes, the traffic redirection command is automatically set up via an init container.

Prerequisites

Kubernetes

  • To use transparent proxy on Kubernetes, Consul-helm >= 0.32.0 and Consul-k8s >= 0.26.0 are required in addition to Consul >= 1.10.0.
  • If the default policy for ACLs is “deny”, then Service Intentions should be set up to allow intended services to connect to each other. Otherwise, all Connect services can talk to all other services.
  • If using Transparent Proxy, all worker nodes within a Kubernetes cluster must have the ip_tables kernel module running, e.g. modprobe ip_tables.

The Kubernetes integration takes care of registering Kubernetes services with Consul, injecting a sidecar proxy, and enabling traffic redirection.

Upgrading to Transparent Proxy

When upgrading from older versions (i.e Consul-k8s < 0.26.0 or Consul-helm < 0.32.0) to Consul-k8s >= 0.26.0 and Consul-helm >= 0.32.0, please make sure to follow the upgrade steps here.

Configuration

Enabling Transparent Proxy

Transparent proxy can be enabled in Kubernetes on the whole cluster via the Helm value:

  1. connectInject:
  2. transparentProxy:
  3. defaultEnabled: true

It can also be enabled on a per namespace basis by setting the label consul.hashicorp.com/transparent-proxy=true on the Kubernetes namespace. This will override the Helm value connectInject.transparentProxy.defaultEnabled and define the default behavior of Pods in the namespace. For example:

  1. kubectl label namespaces my-app "consul.hashicorp.com/transparent-proxy=true"

It can also be enabled on a per service basis via the annotation consul.hashicorp.com/transparent-proxy=true on the Pod for each service, which will override both the Helm value and the namespace label:

  1. apiVersion: v1
  2. kind: Service
  3. metadata:
  4. name: static-server
  5. spec:
  6. selector:
  7. app: static-server
  8. ports:
  9. - protocol: TCP
  10. port: 80
  11. targetPort: 8080
  12. ---
  13. apiVersion: v1
  14. kind: ServiceAccount
  15. metadata:
  16. name: static-server
  17. ---
  18. apiVersion: apps/v1
  19. kind: Deployment
  20. metadata:
  21. name: static-server
  22. spec:
  23. replicas: 1
  24. selector:
  25. matchLabels:
  26. app: static-server
  27. template:
  28. metadata:
  29. name: static-server
  30. labels:
  31. app: static-server
  32. annotations:
  33. 'consul.hashicorp.com/connect-inject': 'true'
  34. 'consul.hashicorp.com/transparent-proxy': '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. serviceAccountName: static-server

Kubernetes HTTP Health Probes Configuration

Traffic redirection interferes with Kubernetes HTTP health probes since the probes expect that kubelet can directly reach the application container on the probe’s endpoint, but that traffic will be redirected through the sidecar proxy, causing errors because kubelet itself is not encrypting that traffic using a mesh proxy. For this reason, Consul allows you to overwrite Kubernetes HTTP health probes to point to the proxy instead. This can be done using the Helm value connectInject.transparentProxy.defaultOverwriteProbes or the Pod annotation consul.hashicorp.com/transparent-proxy-overwrite-probes.

Traffic Redirection Configuration

Pods with transparent proxy enabled will have an init container injected that sets up traffic redirection for all inbound and outbound traffic through the sidecar proxies. This will include all traffic by default, with the ability to configure exceptions on a per-Pod basis. The following Pod annotations allow you to exclude certain traffic from redirection to the sidecar proxies:

Known Limitations

  • Traffic can only be transparently proxied when the address dialed corresponds to the address of a service in the transparent proxy’s datacenter. Services can also dial explicit upstreams in other datacenters without transparent proxy, for example, by adding an annotation such as "consul.hashicorp.com/connect-service-upstreams": "my-service:1234:dc2" to reach an upstream service called my-service in the datacenter dc2.
  • In the deployment configuration where a single Consul datacenter spans multiple Kubernetes clusters, services in one Kubernetes cluster must explicitly dial a service in another Kubernetes cluster using the consul.hashicorp.com/connect-service-upstreams annotation. An example would be "consul.hashicorp.com/connect-service-upstreams": "my-service:1234", where my-service is the service that exists in another Kubernetes cluster and is exposed on port 1234. Although Transparent Proxy is enabled, KubeDNS is not utilized when communicating between services existing on separate Kubernetes clusters.
  • When dialing headless services the request will be proxied using a plain TCP proxy with a 5s connection timeout. Currently the upstream’s protocol and connection timeout are not considered.

Using Transparent Proxy

In Kubernetes, services can reach other services via their KubeDNS address or via Pod IPs, and that traffic will be transparently sent through the proxy. Connect services in Kubernetes are required to have a Kubernetes service selecting the Pods.

Note: In order to use KubeDNS, the Kubernetes service name will need to match the Consul service name. This will be the case by default, unless the service Pods have the annotation consul.hashicorp.com/connect-service overriding the Consul service name.

Transparent proxy is enabled by default in Consul-helm >=0.32.0. The Helm value used to enable/disable transparent proxy for all applications in a Kubernetes cluster is connectInject.transparentProxy.defaultEnabled.

Each Pod for the service will be configured with iptables rules to direct all inbound and outbound traffic through an inbound and outbound listener on the sidecar proxy. The proxy will be configured to know how to route traffic to the appropriate upstream services based on Service Intentions. This means Connect services no longer need to use the consul.hashicorp.com/connect-service-upstreams annotation to configure upstreams explicitly. Once the Service Intentions are set, they can simply address the upstream services using KubeDNS.

As of Consul-k8s >= 0.26.0 and Consul-helm >= 0.32.0, a Kubernetes service that selects application pods is required for Connect applications, i.e:

  1. apiVersion: v1
  2. kind: Service
  3. metadata:
  4. name: sample-app
  5. namespace: default
  6. spec:
  7. selector:
  8. app: sample-app
  9. ports:
  10. - protocol: TCP
  11. port: 80

In the example above, if another service wants to reach sample-app via transparent proxying, it can dial sample-app.default.svc.cluster.local, using KubeDNS. If ACLs with default “deny” policy are enabled, it also needs a ServiceIntention allowing it to talk to sample-app.

Headless Services

For services that are not addressed using a virtual cluster IP, the upstream service must be configured using the DialedDirectly option.

Individual instance addresses can then be discovered using DNS, and dialed through the transparent proxy. When this mode is enabled on the upstream, connect certificates will be presented for mTLS and intentions will be enforced at the destination.

Note that when dialing individual instances HTTP routing rules configured with config entries will not be considered. The transparent proxy acts as a TCP proxy to the original destination IP address.