consul/website/source/docs/guides/connect-production.md

layout	page_title	sidebar_current	description
docs	Connect in Production	docs-guides-connect-production	This guide describes best practices for running Consul Connect in production.

Running Connect in Production

Consul Connect can secure all inter-service communication via mutual TLS. It's designed to work with minimal configuration out of the box, but completing the security checklist and understanding the Consul security model are prerequisites for production deployments.

This guide aims to walk through the steps required to ensure the security guarantees hold.

We assume a cluster is already running with an appropriate number of servers and clients and that other reference material like the deployment and performance guides have been followed.

In practical deployments it may be necessary to incrementally adopt Connect service-by-service. In this case some or all of the advice below may not apply during the transition but should give a good understanding on which security properties have been sacrificed in the interim. The final deployment goal should be to end up compliant with all the advice below.

The steps we need to get to a secure Connect cluster are:

1. Configure ACLs
2. Configure Agent Transport Encryption
3. Bootstrap Certificate Authority
4. Setup Host Firewall
5. Configure Service Instances

Configure ACLs

Consul Connect's security is based on service identity. In practice the identity of the service is only enforcible with sufficiently restrictive ACLs.

This section will not replace reading the full ACL guide but will highlight the specific requirements Connect relies on to ensure it's security properties.

A service's identity, in the form of an x.509 certificate, will only be issued to an API client that has service:write permission for that service. In other words, any client that has permission to register an instance of a service will be able to identify as that service and access all of resources that that service is allowed to access.

A secure ACL setup must meet these criteria:

1. ACL default policy must be deny. It is technically sufficient to keep the default policy of allow but add an explicit ACL denying anonymous service:write. Note however that in this case the Connect intention graph will also default to allow and explicit deny intentions will be needed to restrict service access. Also note that explicit rules to limit who can manage intentions are necessary in this case. It is assumed for the remainder of this guide that ACL policy defaults to deny.
2. Each service must have a distinct ACL token that is restricted to service:write only for the named service. Current Consul ACLs only support prefix matching but in a near-future release we will allow exact name matching. It is possible for all instances of the service to share the same token although best practices is for each instance to get a unique token as described below.

Fine Grained Enforcement

Connect intentions manage access based only on service identity so it is sufficient for ACL tokens to only be unique per service and shared between instances.

It is much better though if ACL tokens are unique per service instance because it limit the blast radius of a compromise.

A future release of Connect will support revoking specific certificates that have been issued. For example if a single node in a datacenter has been compromised, it will be possible to find all certificates issued to the agent on that node and revoke them. This will block all access to the intruder without taking unaffected instances of the service(s) on that node offline too.

While this will work with service-unique tokens, there is nothing stopping an attacker from obtaining certificates while spoofing the agent ID or other identifier – these certificates will not appear to have been issued to the compromised agent and so will not be revoked.

If every service instance has a unique token however, it will be possible to revoke all certificates that were requested under that token. Assuming the attacker can only access the tokens present on the compromised host, this guarantees that any certificate they might have access to or requested directly will be revoked.

In practice, managing per-instance tokens requires automated ACL provisioning, for example using HashiCorp's Vault.

Configure Agent Transport Encryption

Consul's gossip (UDP) and RPC (TCP) communications need to be encrypted otherwise attackers may be able to see tokens and private keys while in flight between the server and client agents or between client agent and application.

Follow the encryption documentation to ensure both gossip encryption and RPC TLS are configured securely.

For now client and server TLS certificates are still managed by manual configuration. In the future we plan to automate more of that with the same mechanisms connect offers to user applications.

Bootstrap Certificate Authority

Consul Connect comes with a built in Certificate Authority (CA) that will bootstrap by default when you first enable Connect on your servers.

To use the built-in CA, enable it in the server's configuration.

connect {
  enabled = true
}

This config change requires a restart which you can perform one server at a time to maintain availability in an existing cluster.

As soon as a server that has Connect enabled becomes the leader, it will bootstrap a new CA and generate it's own private key which is written to the Raft state.

Alternatively, an external private key can be provided via the CA configuration.

External CAs

Connect has been designed with a pluggable CA component so external CAs can be integrated. We will expand the external CA systems that are supported in the future and will allow seamless online migration to a different CA or bootstrapping with an external CA.

For production workloads we recommend using Vault or another external CA once available such that the root key is not stored within Consul state at all.

TODO: link to vault config docs?

Setup Host Firewall

In order to enable inbound connections to connect proxies, you may need to configure host or network firewalls to allow incoming connections to proxy ports.

In addition to Consul agent's communication ports any managed proxies will need to have ports open to accept incoming connections.

Consul will by default assign them ports from a configurable range the default range is 20000 - 20255. If this feature is used, the agent assumes all ports in that range are both free to use (no other processes listening on them) and are exposed in the firewall to accept connections from other service hosts.

Alternatively, managed proxies can have their public ports specified as part of the proxy configuration in the service registration. It is possible to use this exclusively and prevent automated port selection by configuring proxy_min_port and proxy_max_port to both be 0, forcing any managed proxies to have an explicit port configured.

It then becomes the same problem as opening ports necessary for any other application and might be managed by configuration management or a scheduler.

Configure Service Instances

With necessary ACL tokens in place, all service registrations need to have an appropriate ACL token present.

For on-disk configuration the token parameter of the service definition must be set.

For registration via the API the token is passed in the request header or by using the Go client configuration. Note that by default API registration will not allow managed proxies to be configured since it potentially opens a remote execution vulnerability if the agent API endpoints are publicly accessible. This can be configured per-agent.

For examples of service definitions with managed or unmanaged proxies see proxies documentation.

To avoid the overhead of a proxy, applications may natively integrate with connect.

Protect Application Listener

If using any kind of proxy for connect, the application must ensure no untrusted connections can be made to it's unprotected listening port. This is typically done by binding to localhost and only allowing loopback traffic, but may also be achieved using firewall rules or network namespacing.