This guide describes how to run Consul on containers, with Docker as the primary focus. It also describes best practices when running a Consul cluster in production on Docker.
This guide describes critical aspects of operating a Consul cluster that's run inside containers. It primarily focuses on the Docker container runtime, but the principles largely apply to rkt, oci, and other container runtimes as well.
More instructions on how to get started using this image are available at the [official Docker repository page](https://store.docker.com/images/consul)
The container exposes its data directory, `/consul/data`, as a [volume](https://docs.docker.com/engine/tutorials/dockervolumes/). This is where Consul will place its persisted state.
For clients, this stores some information about the cluster and the client's health checks in case the container is restarted. If the volume on a client disappears, it doesn't affect cluster operations.
For servers, this stores the client information plus snapshots and data related to the consensus algorithm and other state like Consul's key/value store and catalog. **Servers need the volume's data to be available when restarting containers to recover from outage scenarios.** Therefore, care must be taken by operators to make sure that volumes containing consul cluster data are not destroyed during container restarts,
When run inside a container, Consul's IP addresses need to be setup properly. Consul has to be configured with an appropriate _cluster address_ as well as a _client address._ In some cases, it might also require configuring an _advertise_address_.
* **Client Address** - The address where other processes on the host contact Consul in order to make HTTP or DNS requests.
* **Advertise Address** - The advertise address is used to change the address that we advertise to other nodes in the cluster. This defaults to the bind address. Consider using this if you use NAT tables in your environment, or in scenarios where you have a routable address that cannot be bound
You will need to tell Consul what its cluster address is when starting so that it binds to the correct interface and advertises a workable interface to the rest of the Consul agents. There are two ways of doing this:
1. Environment Variables: Use the `CONSUL_CLIENT_INTERFACE` and `CONSUL_BIND_INTERFACE` environment variables. In the following example `eth0` is the network interface of the container.
```
docker run \
-d \
-e CONSUL_CLIENT_INTERFACE='eth0' \
-e CONSUL_BIND_INTERFACE='eth0' \
consul agent -server -bootstrap-expect=3
```
2. Address Templates: You can declaratively specify the client and cluster addresses using the formats described in the [go-socketaddr](https://github.com/hashicorp/go-sockaddr) library.
In the following example, the client and bind addresses are declaratively specified for the container network interface 'eth0'
Consul containers can be stopped using the `docker stop <container_id>` command and restarted using `docker start <container_id>`. As long as there are enough servers in the cluster to maintain [quorum](/docs/internals/consensus.html#deployment-table), Consul's [Autopilot](/docs/guides/autopilot.html) feature will handle removing servers whose containers were stopped. Autopilot's default settings are already configured correctly. If you override them, make sure that the following [settings](/docs/agent/options.html#autopilot) are appropriate.
*`server_stabilization_time`should be sufficiently large( on the order of several seconds) so that unstable servers are not added to the cluster until they stabilize.
If a container that was running a leader is stopped, leader election will be triggered causing another server in the cluster to assume leadership.
When a previously stopped server container is restarted using `docker start <container_id>`, and it is configured to obtain a new IP, Autopilot will add it back to the set of Raft peers with the same node-id and the new IP address, after which it can participate as a server again.
**Consul does not currently gracefully handle the situation where all nodes in the cluster running inside a container are restarted at the same time, and they all obtain new IP addresses.** This is because the underlying Raft layer persists the IP address and needs it for leader election operations. Operators must carefully orchestrate restarts of Consul containers that have ephemeral IP addresses to do restarts in small numbers, so that they can gracefully leave the cluster and re-join with their new IP address.
**Snapshot close error** Due to a [known issue](https://github.com/docker/libnetwork/issues/1204) with half close support in Docker, you will see an error message `[ERR] consul: Failed to close snapshot: write tcp <source>-><destination>: write: broken pipe` when saving snapshots. This doesn't affect saving and restoring snapshots when running in Docker.