» DNS Caching

One of the main interfaces to Consul is DNS. Using DNS is a simple way to integrate Consul into an existing infrastructure without any high-touch integration.

By default, Consul serves all DNS results with a 0 TTL value. This prevents any caching. The advantage is that each DNS lookup is always re-evaluated, so the most timely information is served. However, this adds a latency hit for each lookup and can potentially exhaust the query throughput of a cluster.

For this reason, Consul provides a number of tuning parameters that can customize how DNS queries are handled.

» Stale Reads

Stale reads can be used to reduce latency and increase the throughput of DNS queries. The settings used to control stale reads are dns_config.allow_stale, which must be set to enable stale reads, and dns_config.max_stale which limits how stale results are allowed to be.

Since Consul 0.7.1, allow_stale is enabled by default, using a max_stale value that defaults to a near-indefinite threshold (10 years) to allow DNS queries to continue to be served in the event of a long outage with no leader. A new telemetry counter has also been added at consul.dns.stale_queries to track when agents serve DNS queries that are stale by more than 5 seconds.

Doing a stale read allows any Consul server to service a query, but non-leader nodes may return data that is out-of-date. By allowing data to be slightly stale, we get horizontal read scalability. Now any Consul server can service the request, so we increase throughput by the number of servers in a cluster.

If you want to prevent stale reads or limit how stale they can be, you can set allow_stale to false or use a lower value for max_stale. Doing the first will ensure that all reads are serviced by a single leader node. The reads will then be strongly consistent but will be limited by the throughput of a single node.

» Negative Response Caching

Some DNS clients cache negative responses - that is, Consul returning a "not found" style response because a service exists but there are no healthy endpoints. What this means in practice is that cached negative responses may mean that services appear "down" for longer than they are actually unavailable when using DNS for service discovery.

One common example is that Windows will default to caching negative responses for 15 minutes. DNS forwarders may also cache negative responses, with the same effect. To avoid this problem, check the negative response cache defaults for your client operating system and any DNS forwarder on the path between the client and Consul and set the cache values appropriately. In many cases "appropriately" simply is turning negative response caching off to get the best recovery time when a service becomes available again.

» TTL Values

TTL values can be set to allow DNS results to be cached downstream of Consul. Higher TTL values reduce the number of lookups on the Consul servers and speed lookups for clients, at the cost of increasingly stale results. By default, all TTLs are zero, preventing any caching.

To enable caching of node lookups (e.g. "foo.node.consul"), we can set the dns_config.node_ttl value. This can be set to "10s" for example, and all node lookups will serve results with a 10 second TTL.

Service TTLs can be specified in a more granular fashion. You can set TTLs per-service, with a wildcard TTL as the default. This is specified using the dns_config.service_ttl map. The "*" service is the wildcard service.

For example, a dns_config that provides a wildcard TTL and a specific TTL for a service might look like this:

{
  "dns_config": {
    "service_ttl": {
      "*": "5s",
      "web": "30s"
    }
  }
}

This sets all lookups to "web.service.consul" to use a 30 second TTL while lookups to "db.service.consul" or "api.service.consul" will use the 5 second TTL from the wildcard.

Prepared Queries provide an additional level of control over TTL. They allow for the TTL to be defined along with the query, and they can be changed on the fly by updating the query definition. If a TTL is not configured for a prepared query, then it will fall back to the service-specific configuration defined in the Consul agent as described above, and ultimately to 0 if no TTL is configured for the service in the Consul agent.