This topic describes the architecture, components, and resources associated with Consul deployments to Kubernetes. Consul employs the same architectural design on Kubernetes as it does with other platforms (see Architecture), but Kubernetes provides additional benefits that make operating a Consul cluster easier.
Refer to the standard production deployment guide for important information, regardless of the deployment platform.
The server agents are deployed as a
StatefulSet and use persistent volume
claims to store the server state. This also ensures that the
node ID is persisted so that servers
can be rescheduled onto new IP addresses without causing issues. The server agents
are configured with
rules so that they are placed on different nodes. A readiness probe is
configured that marks the pod as ready only when it has established a leader.
Service is registered to represent the servers and exposes ports that are required to communicate to the Consul server pods.
The servers utilize the DNS address of this service to join a Consul cluster, without requiring any other access to the Kubernetes cluster. Additional consul servers may also utilize non-ready endpoints which are published by the Kubernetes service, so that servers can utilize the service for joining during bootstrap and upgrades.
Additionally, a PodDisruptionBudget is configured so the Consul server
cluster maintains quorum during voluntary operational events. The maximum
n is the number of server agents.
Note: Kubernetes and Helm do not delete Persistent Volumes or Persistent Volume Claims when a StatefulSet is deleted, so this must done manually when removing servers.
The client agents are run as a DaemonSet. This places one agent (within its own pod) on each Kubernetes node. The clients expose the Consul HTTP API via a static port (8500) bound to the host port. This enables all other pods on the node to connect to the node-local agent using the host IP that can be retrieved via the Kubernetes downward API. See accessing the Consul HTTP API for an example.
We do not use a NodePort Kubernetes service because requests to node ports get randomly routed to any pod in the service and we need to be able to route directly to the Consul client running on our node.
Note: There is no way to bind to a local-only host port. Therefore, any other node can connect to the agent. This should be considered for security. For a properly production-secured agent with TLS and ACLs, this is safe.
We run Consul clients as a DaemonSet instead of running a client in each application pod as a sidecar because this would turn a pod into a "node" in Consul and also causes an explosion of resource usage since every pod needs a Consul agent. Service registration should be handled via the catalog syncing feature with Services rather than pods.
Note: Due to a limitation of anti-affinity rules with DaemonSets, a client-mode agent runs alongside server-mode agents in Kubernetes. This duplication wastes some resources, but otherwise functions perfectly fine.