proxy

Name

proxy - facilitates both a basic reverse proxy and a robust load balancer.

Description

The proxy has support for multiple backends. The load balancing features include multiple policies, health checks, and failovers. If all hosts fail their health check the proxy plugin will fail back to randomly selecting a target and sending packets to it.

Syntax

In its most basic form, a simple reverse proxy uses this syntax:

proxy FROM TO

FROM is the base domain to match for the request to be proxied.
TO is the destination endpoint to proxy to.

However, advanced features including load balancing can be utilized with an expanded syntax:

proxy FROM TO... {
    policy random|least_conn|round_robin|sequential
    fail_timeout DURATION
    max_fails INTEGER
    health_check PATH:PORT [DURATION]
    except IGNORED_NAMES...
    spray
    protocol [dns [force_tcp]|grpc [insecure|CACERT|KEY CERT|KEY CERT CACERT]]
}

FROM is the name to match for the request to be proxied.
TO is the destination endpoint to proxy to. At least one is required, but multiple may be specified. TO may be an IP:Port pair, or may reference a file in resolv.conf format
policy is the load balancing policy to use; applies only with multiple backends. May be one of random, least_conn, round_robin or sequential. Default is random.
fail_timeout specifies how long to consider a backend as down after it has failed. While it is down, requests will not be routed to that backend. A backend is "down" if CoreDNS fails to communicate with it. The default value is 2 seconds ("2s").
max_fails is the number of failures within fail_timeout that are needed before considering a backend to be down. If 0, the backend will never be marked as down. Default is 1.
health_check will check PATH (on PORT) on each backend. If a backend returns a status code of 200-399, then that backend is marked healthy for double the healthcheck duration. If it doesn't, it is marked as unhealthy and no requests are routed to it. If this option is not provided then health checks are disabled. The default duration is 4 seconds ("4s").
IGNORED_NAMES in except is a space-separated list of domains to exclude from proxying. Requests that match none of these names will be passed through.
spray when all backends are unhealthy, randomly pick one to send the traffic to. (This is a failsafe.)
protocol specifies what protocol to use to speak to an upstream, dns (the default) is plain old DNS. The grpc option will talk to a server that has implemented the DnsService.

Policies

There are four load-balancing policies available:

random (default) - Randomly select a backend
least_conn - Select the backend with the fewest active connections
round_robin - Select the backend in round-robin fashion
sequential - Select the first available backend looking by order of declaration from left to right
first - Deprecated. Use sequential instead

All polices implement randomly spraying packets to backend hosts when no healthy hosts are available. This is to preempt the case where the healthchecking (as a mechanism) fails.

Upstream Protocols

dns : uses the standard DNS exchange. You can pass force_tcp to make sure that the proxied connection is performed over TCP, regardless of the inbound request's protocol.

grpc : extra options are used to control how the TLS connection is made to the gRPC server.

None - No client authentication is used, and the system CAs are used to verify the server certificate.
insecure - TLS is not used, the connection is made in plaintext (not good in production).
CACERT - No client authentication is used, and the file CACERT is used to verify the server certificate.
KEY CERT - Client authentication is used with the specified key/cert pair. The server certificate is verified with the system CAs.
KEY CERT CACERT - Client authentication is used with the specified key/cert pair. The server certificate is verified using the CACERT file.

Metrics

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

coredns_proxy_request_duration_seconds{server, proto, proto_proxy, family, to} - duration per upstream interaction.
coredns_proxy_request_count_total{server, proto, proto_proxy, family, to} - query count per upstream.

Where proxy_proto is the protocol used (dns or grpc) and to is TO specified in the config, proto is the protocol used by the incoming query ("tcp" or "udp"), family the transport family ("1" for IPv4, and "2" for IPv6). Server is the server responsible for the request (and metric). See the documentation in the metrics plugin.

Examples

Proxy all requests within example.org. to a backend system:

proxy example.org 127.0.0.1:9005

Load-balance all requests between three backends (using random policy):

. {
    proxy . 10.0.0.10:53 10.0.0.11:1053 10.0.0.12
}

Same as above, but round-robin style:

. {
    proxy . 10.0.0.10:53 10.0.0.11:1053 10.0.0.12 {
        policy round_robin
    }
}

With health checks and proxy headers to pass hostname, IP, and scheme upstream:

. {
    proxy . 10.0.0.11:53 10.0.0.11:53 10.0.0.12:53 {
        policy round_robin
        health_check /health:8080
    }
}

Proxy everything except requests to miek.nl or example.org

. {
    proxy . 10.0.0.10:1234 {
        except miek.nl example.org
    }
}

Proxy everything except example.org using the host's resolv.conf's nameservers:

. {
    proxy . /etc/resolv.conf {
        except example.org
    }
}

samjegal/proxy