Consul
Intentions in Legacy Mode
1.8.x and earlier: This document only applies in Consul versions 1.8.x and before. If you are using version 1.9.0 or later, refer to the current intentions documentation.
Intentions define access control for service-to-service connections in the service mesh. Intentions can be managed via the API, CLI, or UI.
Intentions are enforced by the proxy or natively integrated application on inbound connections. After verifying the TLS client certificate, the authorize API endpoint is called which verifies the connection is allowed by testing the intentions. If authorize returns false the connection must be terminated.
The default intention behavior is defined by the default ACL policy. If the default ACL policy is "allow all", then all service-to-service connections in the mesh are allowed by default. If the default ACL policy is "deny all", then all service-to-service connections are denied by default.
Intention Basics
Intentions can be managed via the API, CLI, or UI. Please see the respective documentation for each for full details on options, flags, etc. Below is an example of a basic intention to show the basic attributes of an intention. The full data model of an intention can be found in the API documentation.
$ consul intention create -deny web db
Created: web => db (deny)
The intention above is a deny intention with a source of "web" and destination of "db". This says that connections from web to db are not allowed and the connection will be rejected.
When an intention is modified, existing connections will not be affected. This means that changing a connection from "allow" to "deny" today will not kill the connection. Addressing this shortcoming is on the near term roadmap for Consul.
Wildcard Intentions
An intention source or destination may also be the special wildcard
value *
. This matches any value and is used as a catch-all. Example:
$ consul intention create -deny web '*'
Created: web => * (deny)
This example says that the "web" service cannot connect to any service.
Metadata
Arbitrary string key/value data may be associated with intentions. This is unused by Consul but can be used by external systems or for visibility in the UI.
$ consul intention create \
-deny \
-meta description='Hello there' \
web db
...
$ consul intention get web db
Source: web
Destination: db
Action: deny
ID: 31449e02-c787-f7f4-aa92-72b5d9b0d9ec
Meta[description]: Hello there
Created At: Friday, 25-May-18 02:07:51 CEST
Precedence and Match Order
Intentions are matched in an implicit order based on specificity, preferring
deny over allow. Specificity is determined by whether a value is an exact
specified value or is the wildcard value *
.
The full precedence table is shown below and is evaluated
top to bottom, with larger numbers being evaluated first.
Source Namespace | Source Name | Destination Namespace | Destination Name | Precedence |
---|---|---|---|---|
Exact | Exact | Exact | Exact | 9 |
Exact | * | Exact | Exact | 8 |
* | * | Exact | Exact | 7 |
Exact | Exact | Exact | * | 6 |
Exact | * | Exact | * | 5 |
* | * | Exact | * | 4 |
Exact | Exact | * | * | 3 |
Exact | * | * | * | 2 |
* | * | * | * | 1 |
The precedence value can be read from the API after an intention is created. Precedence cannot be manually overridden today. This is a feature that will be added in a later version of Consul.
In the case the two precedence values match, Consul will evaluate intentions based on lexicographical ordering of the destination then source name. In practice, this is a moot point since authorizing a connection has an exact source and destination value so its impossible for two valid non-wildcard intentions to match.
The numbers in the table above are not stable. Their ordering will remain fixed but the actual number values may change in the future.
Consul Enterprise - Namespaces are an Enterprise feature. In Consul OSS any of the rows in
the table with a *
for either the source namespace or destination namespace are not applicable.
Intention Management Permissions
Intention management can be protected by ACLs. Permissions for intentions are destination-oriented, meaning the ACLs for managing intentions are looked up based on the destination value of the intention, not the source.
Intention permissions are by default implicitly granted at read
level
when granting service:read
or service:write
. This is because a
service registered that wants to use service mesh needs intentions:read
for its own service name in order to know whether or not to authorize
connections. The following ACL policy will implicitly grant intentions:read
(note read) for service web
.
service "web" {
policy = "write"
}
It is possible to explicitly specify intention permissions. For example, the following policy will allow a service to be discovered without granting access to read intentions for it.
service "web" {
policy = "read"
intentions = "deny"
}
Note that intentions:read
is required for a token that a mesh-enabled
service uses to register itself or its proxy. If the token used does not
have intentions:read
then the agent will be unable to resolve intentions
for the service and so will not be able to authorize any incoming connections.
Security Note: Explicitly allowing intentions:write
on the token you
provide to a service instance at registration time opens up a significant
additional vulnerability. Although you may trust the service team to define
which inbound connections they accept, using a combined token for registration
allows a compromised instance to to redefine the intentions which allows many
additional attack vectors and may be hard to detect. We strongly recommend only
delegating intentions:write
using tokens that are used by operations teams or
orchestrators rather than spread via application config, or only manage
intentions with management tokens.
Performance and Intention Updates
The intentions for services registered with a Consul agent are cached locally on that agent. They are then updated via a background blocking query against the Consul servers.
Service mesh connection attempts require only local agent communication for authorization and generally only impose microseconds of latency to the connection. All actions in the data path of connections require only local data to ensure minimal performance overhead.
Updates to intentions are propagated nearly instantly to agents since agents maintain a continuous blocking query in the background for intention updates for registered services.
Because all the intention data is cached locally, the agents can fail static. Even if the agents are severed completely from the Consul servers, inbound connection authorization continues to work for a configured amount of time. Changes to intentions will not be picked up until the partition heals, but will then automatically take effect when connectivity is restored.