High availability mode (HA)

Vault supports a multi-server mode for high availability. This mode protects against outages by running multiple Vault servers. High availability mode is automatically enabled when using a data store that supports it.

You can tell if a data store supports high availability mode ("HA") by starting the server and seeing if "(HA available)" is output next to the data store information. If it is, then Vault will automatically use HA mode. This information is also available on the Configuration page.

To be highly available, one of the Vault server nodes grabs a lock within the data store. The successful server node then becomes the active node; all other nodes become standby nodes. At this point, if the standby nodes receive a request, they will either forward the request or redirect the client depending on the current configuration and state of the cluster -- see the sections below for details. Due to this architecture, HA does not enable increased scalability. In general, the bottleneck of Vault is the data store itself, not Vault core. For example: to increase the scalability of Vault with Consul, you would generally scale Consul instead of Vault.

Certain storage backends can support high availability mode, which enable them to store both Vault's information in addition to the HA lock. However, Vault also supports split data/HA mode, whereby the lock value and the rest of the data live separately. This can be done by specifying both the storage and ha_storage stanzas in the configuration file with different backends. For instance, a Vault cluster can be set up to use Consul as the ha_storage to manage the lock, and use Amazon S3 as the storage for all other persisted data.

The sections below explain the server communication patterns and each type of request handling in more detail. At a minimum, the requirements for redirection mode must be met for an HA cluster to work successfully.

Server-to-Server communication

Both methods of request handling rely on the active node advertising information about itself to the other nodes. Rather than over the network, this communication takes place within Vault's encrypted storage; the active node writes this information and unsealed standby Vault nodes can read it.

For the client redirection method, this is the extent of server-to-server communication -- no direct communication with only encrypted entries in the data store used to transfer state.

For the request forwarding method, the servers need direct communication with each other. In order to perform this securely, the active node also advertises, via the encrypted data store entry, a newly-generated private key (ECDSA-P521) and a newly-generated self-signed certificate designated for client and server authentication. Each standby uses the private key and certificate to open a mutually-authenticated TLS 1.2 connection to the active node via the advertised cluster address. When client requests come in, the requests are serialized, sent over this TLS-protected communication channel, and acted upon by the active node. The active node then returns a response to the standby, which sends the response back to the requesting client.

Request forwarding

If request forwarding is enabled (turned on by default in 0.6.2), clients can still force the older/fallback redirection behavior (see below) if desired by setting the X-Vault-No-Request-Forwarding header to any non-empty value.

Successful cluster setup requires a few configuration parameters, although some can be automatically determined.

Client redirection

If X-Vault-No-Request-Forwarding header in the request is set to a non-empty value, the standby nodes will redirect the client using a 307 status code to the active node's redirect address.

This is also the fallback method used when request forwarding is turned off or there is an error performing the forwarding. As such, a redirect address is always required for all HA setups.

Some HA data store drivers can autodetect the redirect address, but it is often necessary to configure it manually via a top-level value in the configuration file. The key for this value is api_addr and the value can also be specified by the VAULT_API_ADDR environment variable, which takes precedence.

What the api_addr value should be set to depends on how Vault is set up. There are two common scenarios: Vault servers accessed directly by clients, and Vault servers accessed via a load balancer.

In both cases, the api_addr should be a full URL including scheme (http/https), not simply an IP address and port.

Direct access

When clients are able to access Vault directly, the api_addr for each node should be that node's address. For instance, if there are two Vault nodes:

  • A, accessed via https://a.vault.mycompany.com:8200

  • B, accessed via https://b.vault.mycompany.com:8200

Then node A would set its api_addr to https://a.vault.mycompany.com:8200 and node B would set its api_addr to https://b.vault.mycompany.com:8200.

This way, when A is the active node, any requests received by node B will cause it to redirect the client to node A's api_addr at https://a.vault.mycompany.com, and vice-versa.

Behind load balancers

Sometimes clients use load balancers as an initial method to access one of the Vault servers, but actually have direct access to each Vault node. In this case, the Vault servers should actually be set up as described in the above section, since for redirection purposes the clients have direct access.

However, if the only access to the Vault servers is via the load balancer, the api_addr on each node should be the same: the address of the load balancer. Clients that reach a standby node will be redirected back to the load balancer; at that point hopefully the load balancer's configuration will have been updated to know the address of the current leader. This can cause a redirect loop and as such is not a recommended setup when it can be avoided.

Per-Node cluster listener addresses

Each listener block in Vault's configuration file contains an address value on which Vault listens for requests. Similarly, each listener block can contain a cluster_address on which Vault listens for server-to-server cluster requests. If this value is not set, its IP address will be automatically set to same as the address value, and its port will be automatically set to the same as the address value plus one (so by default, port 8201).

Note that only active nodes have active listeners. When a node becomes active it will start cluster listeners, and when it becomes standby it will stop them.

Per-Node cluster address

Similar to the api_addr, cluster_addr is the value that each node, if active, should advertise to the standbys to use for server-to-server communications, and lives as a top-level value in the configuration file. On each node, this should be set to a host name or IP address that a standby can use to reach one of that node's cluster_address values set in the listener blocks, including port. (Note that this will always be forced to https since only TLS connections are used between servers.)

This value can also be specified by the VAULT_CLUSTER_ADDR environment variable, which takes precedence.

Storage support

Currently there are several storage backends that support high availability mode, including Consul, ZooKeeper and etcd. These may change over time, and the configuration page should be referenced.

Enclaive recommends Vault Integrated Storage as the default HA backend for new deployments of Vault. Consul Storage Backend is also a supported option and used by many production deployments. See the comparison chart for help deciding which option is best for you.

If you're interested in implementing another backend or adding HA support to another backend, we'd love your contributions. Adding HA support requires implementing the physical.HABackend interface for the storage backend.

Last updated