One of the core concepts when setting up Envoy in production is separating the data plane— the Envoy instances that route your traffic—from the control plane, which acts as the source of truth for the current state of your infrastructure and your desired configuration.
You can start with a static config file as a control plane, but in most cases, teams quickly move from a static config file to an API service. Centralizing configuration in a service makes updates more reliable, taking advantage of Envoy’s ability to query for new configuration and reload it on the fly. The first step in setting up your control plane is connecting your service discovery. This is generally broken down into three steps:
Any control plane should implement the Envoy v2 xDS APIs. For the purpose of publishing service discovery data, you’ll need to implement the Cluster Discovery Service (CDS) and the Endpoint Discovery Service (EDS). Here are several options to avoid starting from scratch:
Rotor is a fast, lightweight xDS implementation with service discovery integration to Kubernetes, Consul, AWS, and more. It provides a set of defaults for routing and listeners (RDS and LDS). It’s also part of the commercial solution Houston, by Turbine Labs, which adds more configuration around routing, resilience, and metrics.
The Envoy repository provides go-control-plane, an open-source stub implementation. If you want to get your hands dirty with exactly how everything is pulled from service discovery, cached locally, and served, this is a great starting point.
If you’re running in Kubernetes, the Istio project has a control plane implementation called Pilot. It takes YAML files and turns them into xDS responses. Don’t be scared by the scope of Istio — Pilot can be used separately to configure Envoy, without pulling in all the other services like Mixer.
In Envoy’s vernacular, a “cluster” is a named group of hosts/ports, over which it will load balance traffic. You may call clusters services, microservices, or APIs. Envoy will periodically poll the CDS endpoint for cluster configuration, and expects a response like:
version_info: "0" resources: - "@type": type.googleapis.com/envoy.api.v2.Cluster name: some_service connect_timeout: 1.0s lb_policy: ROUND_ROBIN type: EDS eds_cluster_config: eds_config: api_config_source: api_type: GRPC cluster_names: [xds_cluster]
Each service collected by your service discovery maps to one item under “resources.” There are a few additional bits of information that are specific to load balancing in Envoy that you’ll need to set:
lb_policy sets the type of load balancing. See what Envoy supports in the
connect_timeout sets the connection timeout. Lower is better: start with 1
second, and raise it if you’re getting flapping due to network latency.
api_type sets the protocol for the service, which your service discovery may
not store directly. Envoy needs to know this to check connections.
It’s possible to hardcode the list of endpoints, though if your infrastructure is dynamic, you’ll want to set “type” to EDS, which tells Envoy to poll the EDS API for a list of available IP/ports.
For the full specifications, see the Envoy docs. Once this is configured, you can populate the endpoints that serve traffic for this cluster.
Envoy defines an “endpoint” as an IP and port available within a cluster. In order to balance traffic across a service, Envoy expects the API to provide a list of endpoints for each service. Envoy periodically polls the EDS endpoint, generating a response:
version_info: “0” resources: - “@type”: type.googleapis.com/envoy.api.v2.ClusterLoadAssignment cluster_name: some_service endpoints: — lb_endpoints: — endpoint: address: socket_address: address: 127.0.0.2 port_value: 1234
This is simpler than defining clusters, because the only thing Envoy needs to know is which cluster(s) this endpoint belongs to.
Envoy treats CDS/EDS service discovery as advisory and eventually consistent; if traffic to an endpoint fails too often, the endpoint is removed from the load balancer until healthy again. There’s no need to aggressively remove endpoints from clusters if they’re unhealthy. Envoy does that for you!
If you have a lot of services, you’ll find that the responses from CDS and EDS are fairly overwhelming. Envoy can handle them, but if something goes wrong, making sense of a 5,000-line API response can be quite challenging. The industry standard is to partition your configs in two ways:
Partition by datacenter / region. In general, services in one datacenter don’t need to know about the exact endpoint available in other datacenters. To set up a trickle of traffic between regions (“backhaul,” making the service robust to region-specific failures), add the remote datacenter’s front proxy to the local load balancer.
Partition by service need. While generally not feasible for an initial roll-out, the most sophisticated Envoy deployments limit intra-service communication by only configuring Envoy sidecars to talk to a whitelist of services. This helps manage the complexity of having 1,000 microservices talk to each other at any time. It also provides some security protection by preventing services from making unexpected calls.
In general, partitioning configuration makes it easier to operate both the Envoy fleet and individual services, at the expense of making the control plane more complex. Since the control plane isn’t in the critical path of customer requests, this tends to be a net win for overall system resilience. Many organizations have reported process wins by making the routing configs (discussed in the next section) partitioned and self-service, as well.
Once your control plane knows about all the available services, it’s time to configure the routes on top of those services. Learn how to set up the Route Discovery Service here.