6.7 KiB
Custom /etc/resolv.conf
- Status: pending
- Version: alpha
- Implementation owner: Bowei Du <bowei@google.com>, Zihong Zheng <zihongz@google.com>
Overview
The /etc/resolv.conf in a pod is managed by Kubelet and its contents are
generated based on pod.dnsPolicy. For dnsPolicy: Default, the search and
nameserver fields are taken from the resolve.conf on the node where the pod
is running. If the dnsPolicy is ClusterFirst, the search contents of the
resolv.conf is the hosts resolv.conf augmented with the following options:
- Search paths to add aliases for domain names in the same namespace and cluster suffix.
options ndotsto 5 to ensure the search paths are searched for all potential matches.
The configuration of both search paths and ndots results in query
amplification of five to ten times for non-cluster internal names. This is due
to the fact that each of the search path expansions must be tried before the
actual result is found. This order of magnitude increase of query rate imposes a
large load on the kube-dns service. At the same time, there are user
applications do not need the convenience of the name aliases and do not wish to
pay this performance cost.
Existing workarounds
The current work around for this problem is to specify an FQDN for name
resolution. Any domain name that ends with a period (e.g. foo.bar.com.) will
not be search path expanded. However, use of FQDNs is not well-known practice
and imposes application-level changes. Cluster operators may not have the luxury
of enforcing such a change to applications that run on their infrastructure.
It is also possible for the user to insert a short shell script snippet that
rewrites resolv.conf on container start-up. This has the same problems as the
previous approach and is also awkward for the user. This also forces the
container to have additional executable code such as a shell or scripting engine
which increases the applications security surface area.
Proposal sketch
This proposal gives users a way to overlay tweaks into the existing
DnsPolicy. A new PodSpec field dnsParams will contains fields that are
merged with the settings currently selected with DnsPolicy.
The fields of DnsParams are:
nameserversis a list of additional nameservers to use for resolution. Onresolv.confplatforms, these are entries tonameserver.searchis a list of additional search path subdomains. Onresolv.confplatforms, these are entries to thesearchsetting. These domains will be appended to the existing search path.optionsthat are an OS-dependent list of (name, value) options. These values are NOT expected to be generally portable across platforms. For containers that use/etc/resolv.confstyle configuration, these correspond to the parameters passed to theoptionlines. Options will override if their names coincide, i.e, if theDnsPolicysetsndots:5andndots:1appears in theSpec, then the final value will bendots:1.
For users that want to completely customize their resolution configuration, we
add a new DnsPolicy: Custom that does not define any settings. This is
essentially an empty resolv.conf with no fields defined.
Pod API examples
Host /etc/resolv.conf
Assume in the examples below that the host has the following /etc/resolv.conf:
nameserver 10.1.1.10
search foo.com
options ndots:1
Override DNS server and search paths
In the example below, the user wishes to use their own DNS resolver and add the pod namespace and a custom expansion to the search path, as they do not use the other name aliases:
# Pod spec
apiVersion: v1
kind: Pod
metadata: {"namespace": "ns1", "name": "example"}
spec:
...
dnsPolicy: Custom
dnsParams:
nameservers: ["1.2.3.4"]
search:
- ns1.svc.cluster.local
- my.dns.search.suffix
options:
- name: ndots
value: 2
- name: edns0
The pod will get the following /etc/resolv.conf:
nameserver 1.2.3.4
search ns1.svc.cluster.local my.dns.search.suffix
options ndots:2 edns0
Overriding ndots
Override ndots:5 in ClusterFirst with ndots:1. This keeps all of the
settings intact:
dnsPolicy: ClusterFirst
dnsParams:
- options:
- name: ndots
- value: 1
Resulting resolv.conf:
nameserver 10.0.0.10
search default.svc.cluster.local svc.cluster.local cluster.local foo.com
options ndots:1
API changes
type PodSpec struct {
...
DNSPolicy string `json:"dnsPolicy,omitempty"`
DNSParams *PodDNSParams `json:"dnsParams,omitempty"`
...
}
type PodDNSParams struct {
Nameservers []string `json:"nameservers,omitempty"`
Search []string `json:"search,omitempty"`
Options []PodDNSParamsOption `json:"options,omitempty" patchStrategy:"merge" patchMergeKey:"name"`
}
type PodDNSParamsOption struct {
Name string `json:"name"`
Value *string `json:"value,omitempty"`
}
Semantics
Let the following be the Go representation of the resolv.conf:
type ResolvConf struct {
Nameserver []string // "nameserver" entries
Search []string // "search" entries
Options []PodDNSParamsOption // "options" entries
}
Let var HostResolvConf ResolvConf be the host resolv.conf.
Then the final Pod resolv.conf will be:
func podResolvConf() ResolvConf {
var podResolv ResolvConf
switch (pod.DNSPolicy) {
case "Default":
podResolv = HostResolvConf
case "ClusterFirst:
podResolv.Nameservers = []string{ KubeDNSClusterIP }
podResolv.Search = ... // populate with ns.svc.suffix, svc.suffix, suffix, host entries...
podResolv.Options = []PodDNSParamsOption{{"ndots","5" }}
case "Custom": // start with empty `resolv.conf`
break
}
// Append the additional nameservers.
podResolv.Nameservers = append(Nameservers, pod.DNSParams.Nameservers...)
// Append the additional search paths.
podResolv.Search = append(Search, pod.DNSParams.Search...)
// Merge the DnsParams.Options with the options derived from the given DNSPolicy.
podResolv.Options = mergeOptions(pod.Options, pod.DNSParams.Options)
return podResolv
}
Invalid configurations
The follow configurations will result in an invalid Pod spec:
- Nameservers or search paths exceed system limits. (Three nameservers, six
search paths, 256 characters for
glibc). - Invalid option appears for the given platform.