Authorization
Warning
Authorization is currently experimental and is not recommended for use in production.
Introduction
Multi-tenant database implies a couple of new concepts that did not previously exist in FoundationDB. The first is the concept of privilege levels: we have data-plane clients whose typical workload is limited to accessing a tenant keyspace. On the other hand, we have control-plane clients or administrators who may read or update cluster-wide configurations through system keyspace. These operations also include creation and deletion of tenants. The second is access control: with multiple tenant keyspaces, it comes naturally that we would want to restrict database access of a client to a subset of them.
Privilege Levels
Authorization feature extends FoundationDB’s existing TLS policy to distinguish administrators from data-plane clients, making TLS configuration a prerequisite for enabling authorization. There are only two privilege levels: trusted versus untrusted clients. Trusted clients are authorized to perform any operation that pre-authorization FoundationDB clients used to perform, including those accessing the system keyspace. Untrusted clients may only request what is necessary to access tenant keyspaces for which they are authorized. Untrusted clients are blocked from accessing anything in the system keyspace or issuing management operations that modify the cluster in any way.
In order to be considered a trusted client, a client needs to be configured with a valid chain of X.509 certificates and a private key,
and its certificate chain must be trusted by the server. In other words, a client must successfully complete a mutual TLS authentication.
Additionally, if the server was configured with trusted IP subnets, i.e. run with one or more --trusted-subnet-SUBNET_NAME
followed by a CIDR block describing the subnet,
then the client’s IP as seen from the server must belong to at least one of the subnets.
Choosing to respond with an empty certificate chain during client authentication marks the client as untrusted. If the server specifies a list of trusted subnets and the client’s server-facing IP is not part of any of the subnets, then the client is untrusted even if it successfully completes a mutual TLS authentication.
Note
Presenting a bad or untrusted certificate chain causes the server to break the client connection and eventually throttle the client. It does not let the client connect untrusted.
Access Control
To restrict untrusted client’s database access to a subset of tenant keyspaces, authorization feature allows database administrators to grant tenant-scoped access in the form of JSON Web Tokens. Token verification is performed against a set of named public keys written in JWK Set format. A token’s header part must contain the key identifier of the public key which shall be used to verify the token itself. Below is the list of token fields recognized by FoundationDB. Note that some of the fields are recognized by FoundationDB but not actively used in enforcing security, pending future implementation. Those fields are marked as NOT required.
Containing Part |
Field Name |
Required |
Purpose |
Reference |
---|---|---|---|---|
Header |
|
Yes |
Type of JSON Web Signature. Must be |
|
Header |
|
Yes |
Algorithm used to generate the signature. Only
|
|
Header |
|
Yes |
Name of public key with which to verify the token.
Must match the |
|
Claim |
|
Yes |
Timestamp after which token is not accepted. |
|
Claim |
|
Yes |
Timestamp before which token is not accepted. |
|
Claim |
|
Yes |
Timestamp at which token was issued. |
|
Claim |
|
Yes |
One or more |
N/A |
Claim |
|
No |
Issuer of the token. |
|
Claim |
|
No |
Subject of the token. |
|
Claim |
|
No |
Intended recipients of the token. Must be an array. |
|
Claim |
|
No |
String that uniquely identifies a token. |
Public keys with which to verify the token must be serialized in JWK Set format and stored in a file.
The location of the key set file must be passed as command line argument --authorization-public-key-file
to the fdbserver
executable.
Public keys in the set must be either RSA public keys
containing n
and e
parameters, each containing Base64urlUInt-encoded modulus and exponent,
or Elliptic Curve public keys on a P-256
curve,
where crv
parameter is set to P-256
and x
and y
parameters contain
base64url-encoded affine coordinates.
In addition, each public key JSON object in set must contain kty
(set to either EC
or RSA
) field to indicate public key algorithm,
along with kid
, and alg
fields to be compared against their token header counterparts.
Private keys are strongly recommended against being included in the public key set and, if found, are excluded from consideration.
Note
By design, FoundationDB authorization feature does not support revocation of outstanding tokens. Use extra caution in signing tokens with long token durations.
Enabling Clients to use Authorization Tokens
In order to use an untrusted client with an authorization token, a client must be configured to trust the server’s CA,
but must not be configured to use the client’s own certificates and keys.
More concretely, the client’s TLS_CA_FILE
must include the server’s root CA certificate,
but the client must not be configured with its own TLS_CERTIFICATE_FILE
or TLS_KEY_FILE
, neither programmatically nor by environment variable.
Before performing a tenant data read or update, a client must set AUTHORIZATION_TOKEN
transaction option with the token string as argument.
It is the client’s responsibility to keep the token up-to-date, by timely assigning a new token to the transaction object.
Note
The TLS authentication mode of an untrusted client is similar to how typical web browsers connect to TLS-enabled web services. They authenticate the server using their bundle of trusted root CA certificates, but they do not authenticate themselves to the server.
Public Key Rotation
FoundationDB’s internal public key set automatically refreshes itself based on the key set file’s latest content every PUBLIC_KEY_FILE_REFRESH_INTERVAL_SECONDS
seconds.
The in-memory set of public keys does not update unless the key file holds a correct JWK Set.
Token Caching
In a single-threaded runtime environment such as FoundationDB, it is important not to let the main thread be overloaded with computationally expensive operations,
such as token signature verification. FoundationDB internally caches the tokens that are considered valid at the time of verification in a fixed-size cache,
whose size may be configured using TOKEN_CACHE_SIZE
knob.
Note
Token cache is independent of the active public key set. Once the token reaches the cache, it is valid until its expiration time, regardless of any key rotation that takes place thereafter.
Allowing Untrusted Clients to Access Tenant Data Without Tokens
Rolling out a public key distribution infrastructure and an authorization-enabled FoundationDB cluster in lockstep might not be feasible with large scale distributed systems.
To support incremental rollout, authorization feature introduces ALLOW_TOKENLESS_TENANT_ACCESS
boolean knob,
which preserves the TLS-based privilege level policy without untrusted clients having to set authorization tokens to their transactions in order to access tenant data.
With this knob active, any authorization token assigned to the client transaction is simply ignored.