Envoy’s cluster manager manages all configured upstream clusters. Just as the Envoy configuration can contain any number of listeners, the configuration can also contain any number of independently configured upstream clusters.
Upstream clusters and hosts are abstracted from the network/HTTP filter stack given that upstream clusters and hosts may be used for any number of different proxy tasks. The cluster manager exposes APIs to the filter stack that allow filters to obtain a L3/L4 connection to an upstream cluster, or a handle to an abstract HTTP connection pool to an upstream cluster (whether the upstream host supports HTTP/1.1 or HTTP/2 is hidden). A filter stage determines whether it needs an L3/L4 connection or a new HTTP stream and the cluster manager handles all of the complexity of knowing which hosts are available and healthy, load balancing, thread local storage of upstream connection data (since most Envoy code is written to be single threaded), upstream connection type (TCP/IP, UDS), upstream protocol where applicable (HTTP/1.1, HTTP/2), etc.
Clusters known to the cluster manager can be configured either statically, or fetched dynamically via the cluster discovery service (CDS) API. Dynamic cluster fetches allow more configuration to be stored in a central configuration server and thus requires fewer Envoy restarts and configuration distribution.
Cluster manager configuration.
When clusters are initialized both at server boot as well as via CDS, they are “warmed.” This means that clusters do not become available until the following operations have taken place.
Initial service discovery load (e.g., DNS resolution, EDS update, etc.).
Initial active health check pass if active health checking is configured. Envoy will send a health check request to each discovered host to determine its initial health status.
The previous items ensure that Envoy has an accurate view of a cluster before it begins using it for traffic serving.
When discussing cluster warming, the cluster “becoming available” means:
For newly added clusters, the cluster will not appear to exist to the rest of Envoy until it has been warmed. I.e., HTTP routes that reference the cluster will result in either a 404 or 503 (depending on configuration).
For updated clusters, the old cluster will continue to exist and serve traffic. When the new cluster has been warmed, it will be atomically swapped with the old cluster such that no traffic interruptions take place.