A new platform for documentation and tutorials is launching soon.
We are migrating Vault documentation into HashiCorp Developer, our new developer experience.
»Replication (Vault Enterprise)
Vault Enterprise 0.7 adds support for multi-datacenter replication. Before using this feature, it is useful to understand the intended use cases, design goals, and high level architecture.
Replication is based on a primary/secondary (1:N) model with asynchronous replication, focusing on high availability for global deployments. The trade-offs made in the design and implementation of replication reflect these high level goals.
Vault replication is based on a number of common use cases:
Multi-Datacenter Deployments: A common challenge is providing Vault to applications across many datacenters in a highly-available manner. Running a single Vault cluster imposes high latency of access for remote clients, availability loss or outages during connectivity failures, and limits scalability.
Backup Sites: Implementing a robust business continuity plan around the loss of a primary datacenter requires the ability to quickly and easily fail to a hot backup site.
Scaling Throughput: Applications that use Vault for Encryption-as-a-Service or cryptographic offload may generate a very high volume of requests for Vault. Replicating keys between multiple clusters allows load to be distributed across additional servers to scale request throughput.
Based on the use cases for Vault Replication, we had a number of design goals for the implementation:
Availability: Global deployments of Vault require high levels of availability, and can tolerate reduced consistency. During full connectivity, replication is nearly real-time between the primary and secondary clusters. Degraded connectivity between a primary and secondary does not impact the primary's ability to service requests, and the secondary will continue to service reads on last-known data.
Conflict Free: Certain replication techniques allow for potential write conflicts to take place. Particularly, any active/active configuration where writes are allowed to multiple sites require a conflict resolution strategy. This varies from techniques that allow for data loss like last-write-wins, or techniques that require manual operator resolution like allowing multiple values per key. We avoid the possibility of conflicts to ensure there is no data loss or manual intervention required.
Transparent to Clients: Vault replication should be transparent to clients of Vault, so that existing thin clients work unmodified. The Vault servers handle the logic of request forwarding to the primary when necessary, and multi-hop routing is performed internally to ensure requests are processed.
Simple to Operate: Operating a replicated cluster should be simple to avoid administrative overhead and potentially introducing security gaps. Setup of replication is very simple, and secondaries can handle being arbitrarily behind the primary, avoiding the need for operator intervention to copy data or snapshot the primary.
The architecture of Vault replication is based on the design goals, focusing on
the intended use cases. When replication is enabled, a cluster is set as either
a primary or secondary. The primary cluster is authoritative, and is the
only cluster allowed to perform actions that write to the underlying data
storage, such as modifying policies or secrets. Secondary clusters can service
all other operations, such as reading secrets or sending data through
transit, and forward any writes to the primary cluster. Disallowing multiple
primaries ensures the cluster is conflict free and has an authoritative state.
The primary cluster uses log shipping to replicate changes to all of the secondaries. This ensures writes are visible globally in near real-time when there is full network connectivity. If a secondary is down or unable to communicate with the primary, writes are not blocked on the primary and reads are still serviced on the secondary. This ensures the availability of Vault. When the secondary is initialized or recovers from degraded connectivity it will automatically reconcile with the primary.
Lastly, clients can speak to any Vault server without a thick client. If a client is communicating with a standby instance, the request is automatically forwarded to an active instance. Secondary clusters will service reads locally and forward any write requests to the primary cluster. The primary cluster is able to service all request types.
An important optimization Vault makes is to avoid replication of tokens or leases between clusters. Policies and secrets are the minority of data managed by Vault and tend to be relatively stable. Tokens and leases are much more dynamic, as they are created and expire rapidly. Keeping tokens and leases locally reduces the amount of data that needs to be replicated, and distributes the work of TTL management between the clusters. The caveat is that clients will need to re-authenticate if they switch the Vault cluster they are communicating with.
It is important to understand the high-level architecture of replication to ensure the trade-offs are appropriate for your use case. The implementation details may be useful for those who are curious or want to understand more about the performance characteristics or failure scenarios.
Using replication requires a storage backend that supports transactional updates, such as Consul. This allows multiple key/value updates to be performed atomically. Replication uses this to maintain a Write-Ahead-Log (WAL) of all updates, so that the key update happens atomically with the WAL entry creation. The WALs are then used to perform log shipping between the Vault clusters. When a secondary is closely synchronized with a primary, Vault directly streams new WALs to be applied, providing near real-time replication. A bounded set of WALs are maintained for the secondaries, and older WALs are garbage collected automatically.
When a secondary is initialized or is too far behind the primary there may not be enough WALs to synchronize. To handle this scenario, Vault maintains a merkle index of the encrypted keys. Any time a key is updated or deleted, the merkle index is updated to reflect the change. When a secondary needs to reconcile with a primary, they compare their merkle indexes to determine which keys are out of sync. The structure of the index allows this to be done very efficiently, usually requiring only two round trips and a small amount of data. The secondary uses this information to reconcile and then switches back into WAL streaming mode.
Performance is an important concern for Vault, so WAL entries are batched and the merkle index is not flushed to disk with every operation. Instead, the index is updated in memory for every operation and asynchronously flushed to disk. As a result, a crash or power loss may cause the merkle index to become out of sync with the underlying keys. Vault uses the ARIES recovery algorithm to ensure the consistency of the index under those failure conditions.
Log shipping traditionally requires the WAL stream to be synchronized, which can introduce additional complexity when a new primary cluster is promoted. Vault uses the merkle index as the source of truth, allowing the WAL streams to be completely distinct and unsynchronized. This simplifies administration of Vault Replication for operators.
»Cluster addresses on the primary
When a cluster is enabled as replication primary, it persists a cluster
definition to storage, under
core/cluster/replicated-dr/info. An optional field of the cluster definition
primary_cluster_addr, which may be provided in the enable request.
Performance standbys regularly issue heartbeat RPC requests to the active node, and
one of the arguments to the RPC is the local node's
primary active node retains these cluster addresses received from its peers in
an in-memory cache named
clusterPeerClusterAddrsCache with a 15s expiry time.
»Cluster addresses on the secondary
When a secondary is enabled, its replication activation token (obtained from
the primary) includes a
primary_cluster_addr field. This is taken from the
persisted cluster definition created when the primary was enabled, or if no
primary_cluster_addr was provided at that time, the token contains the
cluster_addr of the current active node in the primary at the time the
activation token is created.
The secondary persists its own version of the cluster definition to storage,
core/cluster/replicated-dr/info. Here the
primary_cluster_addr field is
the one obtained from the activation token.
The secondary active node regularly issues heartbeat RPC requests to the
primary active node. In response to these, the primary returns a response
which includes a
ClusterAddrs field, comprising the contents of its
clusterPeerClusterAddrsCache plus the current active node's
The secondary uses the response to both to update its in-memory record of
known_primary_cluster_addrs, and in addition it persists them to storage
core/primary-addrs/perf. When this happens
it logs the line
"replication: successful heartbeat", which includes the
ClusterAddrs value obtained in the response.
»Secondary RPC address resolution
gRPC is given a list of target addresses for it to use in performing RPC requests. gRPC will discover that performance standbys can't service most RPCs, and will quickly weed out all but the active node cluster address. If the primary active node changes, gRPC will learn that its address is no longer viable and will automatically fail over to the new active node, assuming it's one of the known target addresses.
The secondary runs a background resolver goroutine that, every few seconds,
builds the gRPC target list of addresses for the primary. Its output is logged
at Trace level as
To build the primary cluster address list, the resolver goroutine normally
known_primary_cluster_addrs with the
primary_cluster_addr in the cluster definition in storage.
There are two exceptions to that normal behaviour: the first time the goroutine
goes through its loop, and when gRPC asks for a forced ResolveNow, which
happens when it's unable to perform RPCs on any of its target addresses. In
both these cases, the resolver goroutine issues a special RemoteResolve RPC to
the primary. This RPC is special because unlike all the other replication
RPCs, it can be serviced by performance standbys as well as the active node. In
either case the node will return the
primary_cluster_addr stored in the
primary's cluster definition, if any, or failing that the current active node's
cluster address. The result of the RemoteResolve call gets included in the
list of target addresses the resolver gives to gRPC for regular RPCs to the
Mismatched Cluster Versions: It is not safe to replicate from a newer version of Vault to an older version. When upgrading replicated clusters, ensure that upstream clusters are always on older version of Vault than downstream clusters. See Upgrading Vault for an example.
Read-After-Write Consistency: All write requests are forwarded from secondaries to the primary cluster in order to avoid potential conflicts. While replication is near real-time, it is not instantaneous, meaning there is a potential for a client to write to a secondary and a subsequent read to return an old value. Secondaries attempt to mask this from an individual client making subsequent requests by stalling write requests until the write is replicated or a timeout is reached (2 seconds). If the timeout is reached, the client will receive a warning. Clients can also take steps to protect against this, see Consistency.
Stale Reads: Secondary clusters service reads based on their locally-replicated data. During normal operation updates from a primary are received in near real-time by secondaries. However, during an outage or network service disruption, replication may stall and secondaries may have stale data. The cluster will automatically recover and reconcile any stale data once the outage has recovered, but reads in the intervening period may receive stale data.