coredns/plugin/traffic

traffic

Name

traffic - handout addresses according to assignments from Envoy's xDS.

Description

The traffic plugin is a balancer that allows traffic steering, weighted responses and draining of clusters. The cluster information is retrieved from a service discovery manager that implements the service discovery protocols from Envoy implements. It connect to the manager using the Aggregated Discovery Service (ADS) protocol.

A Cluster in Envoy is defined as: "A group of logically similar endpoints that Envoy connects to." Each cluster has a name, which traffic extends to be a domain name. See "Naming Clusters" below.

The use case for this plugin is when a cluster has endpoints running in multiple (Kubernetes?) clusters and you need to steer traffic to (or away) from these endpoints, i.e. endpoint A needs to be upgraded, so all traffic to it is drained. Or the entire Kubernetes needs to upgraded, and all endpoints need to be drained from it.

Traffic discovers the endpoints via Envoy's xDS protocol (using ADS). Endpoints and clusters are discovered every 10 seconds. The plugin hands out responses that adhere to these assignments. Only endpoints that are healthy are handed out.

Each DNS response contains a single IP address (or SRV record) that's considered the best one. Traffic will load balance A, AAAA and SRV queries. The TTL on these answer is set to 5s. It will only return successful responses either with an answer or otherwise a NODATA response. Queries for non-existent clusters get a NXDOMAIN, where the minimal TTL is also set to 5s.

When an SRV record is returned an endpoint DNS name is synthesized endpoint-0.<cluster>.<zone> that carries the IP address. Querying for these synthesized names works as well.

The traffic plugin has no notion of draining, drop overload and anything that advanced, it just acts upon assignments. This is means that if a endpoint goes down and traffic has not seen a new assignment yet, it will still include this endpoint address in responses.

Load reporting is not supported for the following reason. A DNS query is done by a resolver. Behind this resolver (which can also cache) there may be many clients that will use this reply. The responding server (CoreDNS) has no idea how many clients use this resolver. So reporting a load of +1 on the CoreDNS side can results in anything from 1 to 1000+ of queries on the endpoint, making the load reporting from traffic highly inaccurate.

Syntax

traffic TO...

This enabled the traffic plugin, with a default node ID of coredns and no TLS.

• TO... are the control plane endpoints to connect to. These must start with grpc://. The port number defaults to 443, if not specified.

The extended syntax is available if you want more control.

traffic TO... {
    node ID
    tls CERT KEY CA
    tls_servername NAME
}

• node ID is how traffic identifies itself to the control plane. This defaults to coredns.

• tls CERT KEY CA define the TLS properties for gRPC connection. If this is omitted an insecure connection is attempted. From 0 to 3 arguments can be provided with the meaning as described below

  • tls - no client authentication is used, and the system CAs are used to verify the server certificate
  • tls CA - no client authentication is used, and the file CA is used to verify the server certificate
  • tls CERT KEY - client authentication is used with the specified cert/key pair. The server certificate is verified with the system CAs.
  • tls CERT KEY CA - client authentication is used with the specified cert/key pair. The server certificate is verified using the specified CA file.

• tls_servername NAME allows you to set a server name in the TLS configuration. This is needed because traffic connects to an IP address, so it can't infer the server name from it.

Naming Clusters

When a cluster is named this usually consists out of a single word, i.e. "cluster-v0", or "web". The traffic plugins uses the name(s) specified in the Server Block to create fully qualified domain names. For example if the Server Block specifies lb.example.org as one of the names, and "cluster-v0" is one of the load balanced cluster, traffic will respond to query asking for cluster-v0.lb.example.org. and the same goes for web; web.lb.example.org.

Metrics

What metrics should we do? If any? Number of clusters? Number of endpoints and health?

Ready

This plugin report readiness to the ready plugin. This will happen after a gRPC stream has been established to the control plane.

Metrics

If monitoring is enabled (via the prometheus plugin) then the following metric are exported:

• coredns_traffic_clusters_tracked{} the number of tracked clusters.

Examples

lb.example.org {
    traffic grpc://127.0.0.1:18000 {
        node test-id
    }
    debug
    log
}

This will load balance any names under lb.example.org using the data from the manager running on localhost on port 18000. The node ID will be test-id and no TLS will be used.

Also See

The following documents provide some background on Envoy's control plane.

• https://github.com/envoyproxy/go-control-plane

• https://blog.christianposta.com/envoy/guidance-for-building-a-control-plane-to-manage-envoy-proxy-based-infrastructure/

• https://github.com/envoyproxy/envoy/blob/442f9fcf21a5f091cec3fe9913ff309e02288659/api/envoy/api/v2/discovery.proto#L63

Bugs

Priority and locality information from ClusterLoadAssignments is not used. Multiple TO addresses is not implemented.

TODO

• credentials (other than TLS) - how/what?
• is the protocol correctly implemented? Should we not have a 10s tick, but wait for responses from the control plane?