Vault is a complex system that has many different pieces. To help both users and developers of Vault build a mental model of how it works, this page documents the system architecture.

Glossary

Before describing the architecture, we provide a glossary of terms to help clarify what is being discussed:

• Storage Backend - A storage backend is responsible for durable storage of encrypted data. Backends are not trusted by Vault and are only expected to provide durability. The storage backend is configured when starting the Vault server.

• Barrier - The barrier is cryptographic steel and concrete around the Vault. All data that flows between Vault and the storage backend passes through the barrier. The barrier ensures that only encrypted data is written out, and that data is verified and decrypted on the way in. Much like a bank vault, the barrier must be "unsealed" before anything inside can be accessed.

• Secrets Engine - A secrets engine is responsible for managing secrets. Simple secrets engines like the "kv" secrets engine simply return the same secret when queried. Some secrets engines support using policies to dynamically generate a secret each time they are queried. This allows for unique secrets to be used which allows Vault to do fine-grained revocation and policy updates. As an example, a MySQL secrets engine could be configured with a "web" policy. When the "web" secret is read, a new MySQL user/password pair will be generated with a limited set of privileges for the web server.

• Audit Device - An audit device is responsible for managing audit logs. Every request to Vault and response from Vault goes through the configured audit devices. This provides a simple way to integrate Vault with multiple audit logging destinations of different types.

• Auth Method - An auth method is used to authenticate users or applications which are connecting to Vault. Once authenticated, the auth method returns the list of applicable policies which should be applied. Vault takes an authenticated user and returns a client token that can be used for future requests. As an example, the userpass auth method uses a username and password to authenticate the user. Alternatively, the github auth method allows users to authenticate via GitHub.

• Client Token - A client token (aka "Vault Token") is conceptually similar to a session cookie on a web site. Once a user authenticates, Vault returns a client token which is used for future requests. The token is used by Vault to verify the identity of the client and to enforce the applicable ACL policies. This token is passed via HTTP headers.

• Secret - A secret is the term for anything returned by Vault which contains confidential or cryptographic material. Not everything returned by Vault is a secret, for example system configuration, status information, or policies are not considered secrets. Secrets always have an associated lease. This means clients cannot assume that the secret contents can be used indefinitely. Vault will revoke a secret at the end of the lease, and an operator may intervene to revoke the secret before the lease is over. This contract between Vault and its clients is critical, as it allows for changes in keys and policies without manual intervention.

• Server - Vault depends on a long-running instance which operates as a server. The Vault server provides an API which clients interact with and manages the interaction between all the secrets engines, ACL enforcement, and secret lease revocation. Having a server based architecture decouples clients from the security keys and policies, enables centralized audit logging and simplifies administration for operators.

High-Level Overview

A very high level overview of Vault looks like this:

Let's begin to break down this picture. There is a clear separation of components that are inside or outside of the security barrier. Only the storage backend and the HTTP API are outside, all other components are inside the barrier.

The storage backend is untrusted and is used to durably store encrypted data. When the Vault server is started, it must be provided with a storage backend so that data is available across restarts. The HTTP API similarly must be started by the Vault server on start so that clients can interact with it.

Once started, the Vault is in a sealed state. Before any operation can be performed on the Vault it must be unsealed. This is done by providing the unseal keys. When the Vault is initialized it generates an encryption key which is used to protect all the data. That key is protected by a master key. By default, Vault uses a technique known as Shamir's secret sharing algorithm to split the master key into 5 shares, any 3 of which are required to reconstruct the master key.

The number of shares and the minimum threshold required can both be specified. Shamir's technique can be disabled, and the master key can be used directly for unsealing. Once Vault retrieves the encryption key, it is able to decrypt the data in the storage backend, and enters the unsealed state. Once unsealed, Vault loads all of the configured audit devices, auth methods, and secrets engines.

The configuration of those audit devices, auth methods, and secrets engines must be stored in Vault since they are security sensitive. Only users with the correct permissions should be able to modify them, meaning they cannot be specified outside of the barrier. By storing them in Vault, any changes to them are protected by the ACL system and tracked by audit logs.

After the Vault is unsealed, requests can be processed from the HTTP API to the Core. The core is used to manage the flow of requests through the system, enforce ACLs, and ensure audit logging is done.

When a client first connects to Vault, it needs to authenticate. Vault provides configurable auth methods providing flexibility in the authentication mechanism used. Human friendly mechanisms such as username/password or GitHub might be used for operators, while applications may use public/private keys or tokens to authenticate. An authentication request flows through core and into an auth method, which determines if the request is valid and returns a list of associated policies.

Policies are just a named ACL rule. For example, the "root" policy is built-in and permits access to all resources. You can create any number of named policies with fine-grained control over paths. Vault operates exclusively in a allowed-access mode, meaning that unless access is explicitly granted via a policy, the action is not allowed. Since a user may have multiple policies associated, an action is allowed if any policy permits it. Policies are stored and managed by an internal policy store. This internal store is manipulated through the system backend, which is always mounted at sys/.

Once authentication takes place and an auth method provides a set of applicable policies, a new client token is generated and managed by the token store. This client token is sent back to the client, and is used to make future requests. This is similar to a cookie sent by a website after a user logs in. The client token may have a lease associated with it depending on the auth method configuration. This means the client token may need to be periodically renewed to avoid invalidation.

Once authenticated, requests are made providing the client token. The token is used to verify the client is authorized and to load the relevant policies. The policies are used to authorize the client request. The request is then routed to the secrets engine, which is processed depending on its type. If the secrets engine returns a secret, the core registers it with the expiration manager and attaches a lease ID. The lease ID is used by clients to renew or revoke their secret. If a client allows the lease to expire, the expiration manager automatically revokes the secret.

The core handles logging of requests and responses to the audit broker, which fans the request out to all the configured audit devices. Outside of the request flow, the core performs certain background activity. Lease management is critical, as it allows expired client tokens or secrets to be revoked automatically. Additionally, Vault handles certain partial failure cases by using write ahead logging with a rollback manager. This is managed transparently within the core and is not user visible.

Getting in Depth

This has been a brief high-level overview of the architecture of Vault. There are more details available for each of the sub-systems.

For other details, either consult the code, ask in IRC or reach out to the mailing list.