Using NodeLocal DNSCache in Kubernetes Clusters

FEATURE STATE: Kubernetes v1.18 [stable]

This page provides an overview of NodeLocal DNSCache feature in Kubernetes.

Before you begin

You need to have a Kubernetes cluster, and the kubectl command-line tool must be configured to communicate with your cluster. It is recommended to run this tutorial on a cluster with at least two nodes that are not acting as control plane hosts. If you do not already have a cluster, you can create one by using minikube or you can use one of these Kubernetes playgrounds:

To check the version, enter kubectl version.

Introduction

NodeLocal DNSCache improves Cluster DNS performance by running a DNS caching agent on cluster nodes as a DaemonSet. In today's architecture, Pods in 'ClusterFirst' DNS mode reach out to a kube-dns serviceIP for DNS queries. This is translated to a kube-dns/CoreDNS endpoint via iptables rules added by kube-proxy. With this new architecture, Pods will reach out to the DNS caching agent running on the same node, thereby avoiding iptables DNAT rules and connection tracking. The local caching agent will query kube-dns service for cache misses of cluster hostnames ("cluster.local" suffix by default).

Motivation

  • With the current DNS architecture, it is possible that Pods with the highest DNS QPS have to reach out to a different node, if there is no local kube-dns/CoreDNS instance. Having a local cache will help improve the latency in such scenarios.

  • Skipping iptables DNAT and connection tracking will help reduce conntrack races and avoid UDP DNS entries filling up conntrack table.

  • Connections from the local caching agent to kube-dns service can be upgraded to TCP. TCP conntrack entries will be removed on connection close in contrast with UDP entries that have to timeout (default nf_conntrack_udp_timeout is 30 seconds)

  • Upgrading DNS queries from UDP to TCP would reduce tail latency attributed to dropped UDP packets and DNS timeouts usually up to 30s (3 retries + 10s timeout). Since the nodelocal cache listens for UDP DNS queries, applications don't need to be changed.

  • Metrics & visibility into DNS requests at a node level.

  • Negative caching can be re-enabled, thereby reducing the number of queries for the kube-dns service.

Architecture Diagram

This is the path followed by DNS Queries after NodeLocal DNSCache is enabled:

NodeLocal DNSCache flow

Nodelocal DNSCache flow

This image shows how NodeLocal DNSCache handles DNS queries.

Configuration

This feature can be enabled using the following steps:

  • Prepare a manifest similar to the sample nodelocaldns.yaml and save it as nodelocaldns.yaml.

  • If using IPv6, the CoreDNS configuration file needs to enclose all the IPv6 addresses into square brackets if used in 'IP:Port' format. If you are using the sample manifest from the previous point, this will require you to modify the configuration line L70 like this: "health [__PILLAR__LOCAL__DNS__]:8080"

  • Substitute the variables in the manifest with the right values:

    kubedns=`kubectl get svc kube-dns -n kube-system -o jsonpath={.spec.clusterIP}`
    domain=<cluster-domain>
    localdns=<node-local-address>
    

    <cluster-domain> is "cluster.local" by default. <node-local-address> is the local listen IP address chosen for NodeLocal DNSCache.

    • If kube-proxy is running in IPTABLES mode:

      sed -i "s/__PILLAR__LOCAL__DNS__/$localdns/g; s/__PILLAR__DNS__DOMAIN__/$domain/g; s/__PILLAR__DNS__SERVER__/$kubedns/g" nodelocaldns.yaml
      

      __PILLAR__CLUSTER__DNS__ and __PILLAR__UPSTREAM__SERVERS__ will be populated by the node-local-dns pods. In this mode, the node-local-dns pods listen on both the kube-dns service IP as well as <node-local-address>, so pods can look up DNS records using either IP address.

    • If kube-proxy is running in IPVS mode:

      sed -i "s/__PILLAR__LOCAL__DNS__/$localdns/g; s/__PILLAR__DNS__DOMAIN__/$domain/g; s/,__PILLAR__DNS__SERVER__//g; s/__PILLAR__CLUSTER__DNS__/$kubedns/g" nodelocaldns.yaml
      

      In this mode, the node-local-dns pods listen only on <node-local-address>. The node-local-dns interface cannot bind the kube-dns cluster IP since the interface used for IPVS loadbalancing already uses this address. __PILLAR__UPSTREAM__SERVERS__ will be populated by the node-local-dns pods.

  • Run kubectl create -f nodelocaldns.yaml

  • If using kube-proxy in IPVS mode, --cluster-dns flag to kubelet needs to be modified to use <node-local-address> that NodeLocal DNSCache is listening on. Otherwise, there is no need to modify the value of the --cluster-dns flag, since NodeLocal DNSCache listens on both the kube-dns service IP as well as <node-local-address>.

Once enabled, the node-local-dns Pods will run in the kube-system namespace on each of the cluster nodes. This Pod runs CoreDNS in cache mode, so all CoreDNS metrics exposed by the different plugins will be available on a per-node basis.

You can disable this feature by removing the DaemonSet, using kubectl delete -f <manifest>. You should also revert any changes you made to the kubelet configuration.

StubDomains and Upstream server Configuration

StubDomains and upstream servers specified in the kube-dns ConfigMap in the kube-system namespace are automatically picked up by node-local-dns pods. The ConfigMap contents need to follow the format shown in the example. The node-local-dns ConfigMap can also be modified directly with the stubDomain configuration in the Corefile format. Some cloud providers might not allow modifying node-local-dns ConfigMap directly. In those cases, the kube-dns ConfigMap can be updated.

Setting memory limits

The node-local-dns Pods use memory for storing cache entries and processing queries. Since they do not watch Kubernetes objects, the cluster size or the number of Services / EndpointSlices do not directly affect memory usage. Memory usage is influenced by the DNS query pattern. From CoreDNS docs,

The default cache size is 10000 entries, which uses about 30 MB when completely filled.

This would be the memory usage for each server block (if the cache gets completely filled). Memory usage can be reduced by specifying smaller cache sizes.

The number of concurrent queries is linked to the memory demand, because each extra goroutine used for handling a query requires an amount of memory. You can set an upper limit using the max_concurrent option in the forward plugin.

If a node-local-dns Pod attempts to use more memory than is available (because of total system resources, or because of a configured resource limit), the operating system may shut down that pod's container. If this happens, the container that is terminated (“OOMKilled”) does not clean up the custom packet filtering rules that it previously added during startup. The node-local-dns container should get restarted (since managed as part of a DaemonSet), but this will lead to a brief DNS downtime each time that the container fails: the packet filtering rules direct DNS queries to a local Pod that is unhealthy.

You can determine a suitable memory limit by running node-local-dns pods without a limit and measuring the peak usage. You can also set up and use a VerticalPodAutoscaler in recommender mode, and then check its recommendations.

Last modified August 25, 2022 at 10:05 PM PST: Favor EndpointSlice over Endpoints (1eef742465)