DNS

From Wikitech
This page may be outdated or contain incorrect details. Please update it if you can.

This page describes Wikimedia's DNS setup. Wikimedia use two separate kinds of DNS servers, authoritative nameservers (that respond to queries from third party nameservers for our domains) and recursive resolvers (that resolve DNS queries when any of our servers need to look up a name)

Need to make changes to Wikimedia zones? See #How to make DNS changes. Questions? Please reach out to the Traffic team: #wikimedia-traffic on Libera.

Authoritative nameservers

In Wikimedia's DNS setup, Wikimedia has 3 authoritative public DNS server Hostnames/IPs, which are referenced by the NS records in all of our zones. These service IPs are routed by manual configuration in our Juniper routers to any of the 14 available target production hosts whose role contains profile::dns::auth (dns[1-9][0-9]{3}).

The three nameserver IPs and their current routing are:

  • ns0.wikimedia.org - 208.80.154.238 (currently hosted on dns100[4-6].wikimedia.org in eqiad)
  • ns1.wikimedia.org - 208.80.153.231 (currently hosted on dns200[4-6].wikimedia.org in codfw)
  • ns2.wikimedia.org - 198.35.27.27 (anycasted, announced from all dns hosts)

All 14 servers which run the authdns setup are picked up by the cumin alias A:dns-auth (as opposed to A:dns-rec, which picks up the subset which are also recursors). If you wish to query the authdns service on the individual underlying hosts (as opposed to querying the 3 official service IPs, which could be routed to any of them), you'll need to query the non-standard port 5353, as the authdns service doesn't listen on the host IP's port 53 (depending on which subset of hosts, either nothing listens there, or a recursive dns service listens there).

The servers are running gdnsd with the geoip plugin which is responsible for geographic DNS.

Zonefiles and other configuration are replicated through the use of git fetch and git merge in a set of update scripts. In case of emergency the servers can be synced from any other as well.

All configuration files can be found in /etc/gdnsd/ on all three hosts, with a separate conf file used just for syncing the zonefiles and such, in /etc/wikimedia-authdns.conf.

The main gdns configuration file is /etc/gdnsd/config. It is generated from files in our operations/dns git repo, as are the zone files.

Debugging Public Nameserver IP mapping

All of our authservers support the NSID EDNS option and are configured to report their internal hostnames in response to NSID queries.

The dig or kdig tools can be used both from the public Internet as well as internal hosts to check which internal authdns server is responding to a request to one of the public authdns IPs by using the +nsid option.

Examples (remove the grep part to see the full output):

bblack@cumin1001:~$ dig +nsid @ns0.wikimedia.org en.wikipedia.org A|grep NSID
; NSID: 64 6e 73 31 30 30 31 ("dns1001")
bblack@cumin1001:~$ dig +nsid @ns1.wikimedia.org en.wikipedia.org A|grep NSID
; NSID: 64 6e 73 32 30 30 31 ("dns2001")
bblack@cumin1001:~$ dig +nsid @ns2.wikimedia.org en.wikipedia.org A|grep NSID
; NSID: 64 6e 73 33 30 30 31 ("dns3001")

# Using kdig from an outside host, for which the routers hash ns2 to dns3002 instead:
blblack@dallas:~$ kdig +nsid @ns2.wikimedia.org en.wikipedia.org A|grep NSID
;; NSID: 646E7333303032 "dns3002"

Domain templates

The zone templates are (regular) files in the operations/dns git repo in the templates/ directory.

Each regular file in this directory corresponds to a zone with the same name. Each symbolic link to a regular file in this directory corresponds to a domain alias. So, in this example:

# ls -l templates/mediawiki*
lrwxrwxrwx    1 root     root           13 Jun 19 15:52 templates/mediawiki.com -> mediawiki.org
lrwxrwxrwx    1 root     root           13 Jun 19 15:52 templates/mediawiki.net -> mediawiki.org
-rw-r--r--    1 root     root         1500 Jun 19 15:12 templates/mediawiki.org

...one zone mediawiki.org is listed, with two alias zones, mediawiki.com and mediawiki.net.

The zonefiles generally follow the standard DNS Master File format (aka BIND zone files) specified by section 5 of RFC1035. gdnsd itself and our jinja-based templating add a few other capabilities on top, notably:

Variables and macros

Within the zone template, a few predefined variables and macros can be used, that will be substituted when the actual zonefiles are generated from the template. These include:

$INCLUDE filename origin
Includes another file, which should be located in a subdirectory
@Z
Is replaced by the actual zone name (FQDN) of the zone (works for symlinks, and also include files)
@F
Is replaced by the original origin at the start of the file, same as @Z in the case of the main zonefile, but can be different in an include file
{{ serial_num }}
Causes an SOA serial number to autogenerated in this place approximating a datestamp of last change, derived from git commit history
{{ serial_comment }}
Emits a text suitable only for a comment line, showing the git hash and the first line of the commit message
{{ langlist(...) }}
A list of language subdomain CNAMEs, i.e. a list of all language abbreviations for all languages any Wikimedia project has, generated from helpers/langlist.tmpl.

Other generic jinja2 templating constructs can be used as well, e.g.:

{% for i in range(1, 10) %}
asw{{ i }}-pmtpa	1H	IN A		10.0.1.{{ i }}
{%- endfor %}

Note that if the range is (a,b) then the first entry will be for a but the last entry will be for b-1.

authdns-update

/usr/local/sbin/authdns-update is a simple shell script installed on the authdns servers (dnsXXXX.wikimedia.org), that automates the invocations of the scripts above. It goes through the following steps on each of ns0-2 (needs updating, too lazy to fix completely right now):

  1. ssh to the host
  2. git pull the templates from operations/git repo via authdns-git-pull
  3. generate the zone files from the zone template files via authdns-gen-zones
  4. update the gdnsd config files from the local git repo just updated
  5. sanity checks and reload of the gdns daemon

Basically, authdns-update takes care of everything after you've edited and merged the zonefiles.

authdns-local-update

/usr/sbin/authdns-local-update is used on any of the servers for pulling in updates from any other (presumably up to date) dns server. It can be used to bring a server back up to date after e.g. downtime or a software install/update. It is also used in this way by puppet during initial setup.

Geographic DNS

Geographic DNS makes sure that clients end up using the Wikimedia cluster closest to them, by varying DNS responses based on the (country of the) resolver IP querying. This is handled by the gdns geoip plugin. The config file is in config-geo in the operations/dns repo. Our geoip setup makes use of /usr/share/GeoIP/GeoIPCity.dat (ipv4) and /usr/share/GeoIP/GeoIPv6.dat (ipv6). These are pulled from the volatile directory on the puppet master which is updated regularly by cron. See the geoip module for more information.

How to make DNS changes

This section briefly explains how to do the most common DNS changes.

Change GeoDNS / Depool a Site

When editing this, please also update README.md.

For example, when a certain cluster is down/unreachable, and you want to move all traffic to the others.

Depooling any site:

  • In the operations/dns repo, edit the admin_state file to include something like geoip/generic-map/eqiad => DOWN (see examples within), commit, and run authdns-update from any of the dns servers.

Something to consider during peak traffic times is to watch our peering/transit links for saturation! While link saturation is a paging alert, LibreNMS's view of peering and transit links can help.

Changing records in a zonefile

  • This is handled via the git repo operations/dns
  • Edit the template file templates/zonename locally and check into git, and git review (for gerrit review)
  • Merge your change in gerrit, then login to dns1004.wikimedia.org (or any DNS host in the cumin alias A:dns-rec), and run sudo -i authdns-update. This will pull from operations/dns and generate zonefiles and gnsd configs on each nameserver.
  • This no longer requires you to forward your own key, the systems are set up with their own trusted keys for the sync.
  • Please make sure that the update runs correctly on all dns servers, contact other SREs in case this doesn't happen and you don't know why.
  • Once the script completes, its a good habit to query all three DNS servers to make sure your change has been correctly deployed
    • for example:
for i in 0 1 2; do
   ns=ns${i}.wikimedia.org
   echo $ns
   dig +short @${ns} -t srv _etcd-server-ssl._tcp.dse-k8s-etcd.eqiad.wmnet
done
  • If any auth DNS server failed to response, please contact other SREs and check what's wrong before doing anything else.

Netbox-driven dynamic data

Most of the infrastructure records are dynamically generated via Netbox-driven data. See DNS/Netbox.

Adding a new zone

  1. First, decide if this new zone will use a new, independent zonefile, or will be an alias of another zone
    independent zonefile
    Create the new zone template in the operations/dns repository as templates/zonename. (Copy an existing, relatively clean zonefile like wiktionary.org to start with).
    zone alias
    Make a symbolic link templates/aliasname for the alias to the zone being aliased.
  2. git add the file in, commit, and review on gerrit.
  3. Run authdns-update on ns0 (or any nameserver).
  4. Query all three ns servers to verify that your change took correctly.
  5. If any auth DNS server failed to response, restart it with /etc/init.d/pdns restart

Removing a zone

  • git rm the appropriate file, and merge on gerrit.
  • Log into NS0 and run authdns-update

Adding a new (language) wiki

  1. Add the language code to templates/helpers/langs.tmpl in the operations/dns repo and merge the commit
  2. Normal deployment process will not create the expected results. You must follow the workaround documented at https://phabricator.wikimedia.org/T97051#1994679 for now!

If a certain nameserver is unreachable

When a certain nameserver is unreachable, the others can still be updated from any of the other servers, by running authdns-update there. To skip the unreachable server in the update process, use:

# authdns-update -s "server list"

where server list is a space separated list of FQDNs. Do not forget the quotes, the script will only accept one argument behind -s.

  1. query all three ns servers to verify that your change took correctly.
  2. If any auth DNS server failed to response, restart it with /etc/init.d/pdns restart

Linting the zone files

  • Most scripts here expect to be run from the root directory of a clone of the operations/dns repository.
  • All of the python scripts require Python 3.5+ (e.g. stretch python3)

To run locally the same checks of CI:

  • Partial local CI: install the tox python module and run tox -- -n
  • Complete local CI: install tox python module and the latest WMF debian package of gdnsd and run: tox

For more detailed instructions see the utils/README file inside the repository.

For more details on the WMF-specific DNS zone files consistency check validator script for the internal zone files (wmnet and pointer zones), see the docstring at the top of the file.

Know to which DC a specific IP is redirected

bblack@dns1001:~$ gdnsd_geoip_test
[starts an interactive shell, then input mapname followed by IP like:]
> generic-map 1.2.3.4
generic-map => 1.2.3.4/24 => eqiad, codfw, ulsfo, esams, eqsin

If you just want to do a single lookup you can put the mapname and IP on the commandline too, but if you're doing a bunch, using the interactive shell prevents expensively reloading maps every time

Know which IP the AuthDNS is seeing a query from

For scenarios where the network you're currently connected is not being redirected to the proper DC, while both your IP and your resolver's IP maps to the proper DC.

Lookup reflect.wikimedia.org, it will tell you how the authdns really sees the client or recursor IP.

Update DNS if gerrit or DNS are down (on an emergency only)

There may be some time in which GIT may be unavailable, and some changes may be needed on dns to unbreak it (e.g. for a failover), or that cannot wait for it to be up.

Try to contact someone on SRE/Traffic before trying this, as the workflow is not yet suited for doing manual changes. Also, soon we will have more than 3 authservers, making any manual "ssh to all the authservers and edit files manually" bad. Make sure this is the last and only resort before doing it- dependencies between services make this an unsafe operation.
  1. ssh into every authserver (dns and ips are at DNS#Authoritative_nameservers)
  2. edit the zone data directly in /etc/gdnsd/zones to fix the problem
  3. execute gdnsdctl reload-zones
  4. Remember to repeat it for each authserver, so they are all in sync

Make sure to log and declare that nobody should touch the DNS gerrit or authdns-update, until things are made sane again, otherwise things will break (even more).

Recursive Resolvers

pdns_recursor is run on the following nodes:

  • dns[1001-1002] in eqiad
  • dns[2001-2002] in codfw
  • dns[3001-3002] in esams
  • dns[4001-4002] in ulsfo
  • dns[5001-5002] in eqsin

These use the role::dnsbox role.

The recursive cache software used here is PowerDNS recursor (aka pdns-recursor). These machines also host a local authdns instance and site NTP servers.

Statistics are available at: Grafana DNS Recursor Stats

Wipe caches

To wipe records from the recursors caches to clear wrong records, negative caches, etc, you can run the sre.dns.wipe-cache cookbook:

$ sudo cookbook sre.dns.wipe-cache -h
usage: cookbook [GLOBAL_ARGS] sre.dns.wipe-cache [-h] [-s {eqiad,codfw,esams,ulsfo,eqsin,drmrs}] domains [domains ...]

Cookbook to wipedns cache entries.

    Perform the actions outlined in the following wiki:
    https://wikitech.wikimedia.org/wiki/DNS#Wipe_caches

    Usage example:
        cookbook sre.dns.wipe-cache puppet.esqin.wmnet
        cookbook sre.dns.wipe-cache --site esqin puppet.esqin.wmnet
        cookbook sre.dns.wipe-cache ulsfo.wmnet$
        cookbook sre.dns.wipe-cache host1001.eqiad.wmnet host2001.codfw.wmnet


positional arguments:
  domains               The DNS domains to wipe from the cache. Multiple domains can be passed. The domains can be
                        suffixed with a ‘$’. to delete the whole tree from the cache.

optional arguments:
  -h, --help            show this help message and exit
  -s {eqiad,codfw,esams,ulsfo,eqsin,drmrs}, --site {eqiad,codfw,esams,ulsfo,eqsin,drmrs}
                        This is used to limit the recursors that will be wiped. Passing eqiad here will only wipe dns
                        entries from the eqiad servers (default: None)

How to manually remove, a record from the DNS resolver caches (flush / wipe / clear dns resolver cache)

If you have just added or updated a DNS record on the authoritative nameservers, it may still be cached on the (unrelated) DNS resolvers used by our servers. To clear a record from the resolver cache, use:

# rec_control wipe-cache record-name

on all the DNS resolvers. This will also clear any negative cache records. If you need to clear a PTR record, be sure to use the actual record name, e.g.

 # rec_control wipe-cache 122.36.64.10.in-addr.arpa.

(with the trailing '.').

All of the recursors can be targeted via this cumin command:

 $ sudo cumin 'A:dns-rec' 'rec_control wipe-cache <FQDN.of.server>'

How to manually remove an entire domain from the DNS resolver caches

In larger migrations it might be useful to be able to wipe the cache for an entire domain. For domains for which we are authoritative the impact on the resolvers is minimal given that they query a local instance of gdnsd. Care should be used before wiping the cache for an external domain, as it might affect the time it will take to repopulate the cache from the external DNS authoritative resolvers.

To wipe an entire domain the $ suffix can be used as in this example:

$ sudo cumin 'A:dns-rec' 'rec_control wipe-cache eqsin.wmnet$'

For reverse records it can be expanded in the reverse way, for example the following command will wipe all reverse IPv4 records of the form 10.64.*.*:

$ sudo cumin 'A:dns-rec' 'rec_control wipe-cache 64.10.in-addr.arpa$'

For more information see the official documentation at: https://doc.powerdns.com/recursor/manpages/rec_control.1.html.

See also

  • Portal:Cloud_VPS/Admin/DNS -- documentation on Wikimedia Cloud Services DNS (wikimedia.cloud, wikimediacloud.org, toolforge.org, etc).


External links