Service discovery

When an upstream cluster is defined in the configuration, Envoy needs to know how to resolve the members of the cluster. This is known as service discovery.

Supported service discovery types

Static

Static is the simplest service discovery type. The configuration explicitly specifies the resolved network name (IP address/port, unix domain socket, etc.) of each upstream host.

Strict DNS

When using strict DNS service discovery, Envoy will continuously and asynchronously resolve the specified DNS targets. Each returned IP address in the DNS result will be considered an explicit host in the upstream cluster. This means that if the query returns three IP addresses, Envoy will assume the cluster has three hosts, and all three should be load balanced to. If a host is removed from the result Envoy assumes it no longer exists and will drain traffic from any existing connection pools. Note that Envoy never synchronously resolves DNS in the forwarding path. At the expense of eventual consistency, there is never a worry of blocking on a long running DNS query.

Logical DNS

Logical DNS uses a similar asynchronous resolution mechanism to strict DNS. However, instead of strictly taking the results of the DNS query and assuming that they comprise the entire upstream cluster, a logical DNS cluster only uses the first IP address returned when a new connection needs to be initiated. Thus, a single logical connection pool may contain physical connections to a variety of different upstream hosts. Connections are never drained. This service discovery type is optimal for large scale web services that must be accessed via DNS. Such services typically use round robin DNS to return many different IP addresses. Typically a different result is returned for each query. If strict DNS were used in this scenario, Envoy would assume that the cluster’s members were changing during every resolution interval which would lead to draining connection pools, connection cycling, etc. Instead, with logical DNS, connections stay alive until they get cycled. When interacting with large scale web services, this is the best of all possible worlds: asynchronous/eventually consistent DNS resolution, long lived connections, and zero blocking in the forwarding path.

Original destination

Original destination cluster can be used when incoming connections are redirected to Envoy either via an iptables REDIRECT or TPROXY target or with Proxy Protocol. In these cases requests routed to an original destination cluster are forwarded to upstream hosts as addressed by the redirection metadata, without any explicit host configuration or upstream host discovery. Connections to upstream hosts are pooled and unused hosts are flushed out when they have been idle longer than cleanup_interval_ms, which defaults to 5000ms. If the original destination address is is not available, no upstream connection is opened. Envoy can also pickup the original destination from a HTTP header. Original destination service discovery must be used with the original destination load balancer.

Service discovery service (SDS)

The service discovery service is a generic REST based API used by Envoy to fetch cluster members. Lyft provides a reference implementation via the Python discovery service. That implementation uses AWS DynamoDB as the backing store, however the API is simple enough that it could easily be implemented on top of a variety of different backing stores. For each SDS cluster, Envoy will periodically fetch the cluster members from the discovery service. SDS is the preferred service discovery mechanism for a few reasons:

  • Envoy has explicit knowledge of each upstream host (vs. routing through a DNS resolved load balancer) and can make more intelligent load balancing decisions.
  • Extra attributes carried in the discovery API response for each host inform Envoy of the host’s load balancing weight, canary status, zone, etc. These additional attributes are used globally by the Envoy mesh during load balancing, statistic gathering, etc.

Generally active health checking is used in conjunction with the eventually consistent service discovery service data to making load balancing and routing decisions. This is discussed further in the following section.

On eventually consistent service discovery

Many existing RPC systems treat service discovery as a fully consistent process. To this end, they use fully consistent leader election backing stores such as Zookeeper, etcd, Consul, etc. Our experience has been that operating these backing stores at scale is painful.

Envoy was designed from the beginning with the idea that service discovery does not require full consistency. Instead, Envoy assumes that hosts come and go from the mesh in an eventually consistent way. Our recommended way of deploying a service to service Envoy mesh configuration uses eventually consistent service discovery along with active health checking (Envoy explicitly health checking upstream cluster members) to determine cluster health. This paradigm has a number of benefits:

  • All health decisions are fully distributed. Thus, network partitions are gracefully handled (whether the application gracefully handles the partition is a different story).
  • When health checking is configured for an upstream cluster, Envoy uses a 2x2 matrix to determine whether to route to a host:
Discovery StatusHC OKHC Failed
DiscoveredRouteDon’t Route
AbsentRouteDon’t Route / Delete

Host discovered / health check OK

Envoy will route to the target host.

Host absent / health check OK:

Envoy will route to the target host. This is very important since the design assumes that the discovery service can fail at any time. If a host continues to pass health check even after becoming absent from the discovery data, Envoy will still route. Although it would be impossible to add new hosts in this scenario, existing hosts will continue to operate normally. When the discovery service is operating normally again the data will eventually re-converge.

Host discovered / health check FAIL

Envoy will not route to the target host. Health check data is assumed to be more accurate than discovery data.

Host absent / health check FAIL

Envoy will not route and will delete the target host. This is the only state in which Envoy will purge host data.