Anycast

From Wikitech

Anycast-based Wikimedia deployments include Anycast authoritative DNS, Wikimedia DNS and its accompanying verification server, Durum. In production there is Anycast recursive DNS and the #Syslog. Since October 2023, the NTP address for the Debian installer, ntp.anycast.wmnet is also anycasted.

Anycast works thusly:

  • The VIP (virtual IP) is configured on the servers loopback
  • Bird (routing daemon) advertises the VIP to the routers using BGP
  • (optional) A BFD session is established between Bird and the routers to ensure fast failover in case of server or link failure
  • Anycast_healthchecker monitors the local (anycasted) service by querying it every second
  • If a service failure is detected, the VIP stops being advertised to the routers
  • When the service is restored, anycast_healthchecker waits 10s before re-advertising the IP to avoid flaps
  • The bird service is linked (systemd bind) to the anycast_healthchecker service so bird is stopped if anycast_healthchecker is not running/crashed
  • Time between a local service failure and clients to be redirected to a different server (advertising the same VIP) is 1s max
  • All servers advertise the same VIP worldwide, clients will be be routed to the closest (in the BGP definition) server (same DC, then shorter AS path, etc...) but is not based on latency
  • Routers do per flow load balancing (ECMP) between all local (same site) servers. Hashing is done on IP and port (L4)
  • As last hope backup, in case all servers stop advertising the VIP (eg. global missconfiguration), eqiad and codfw routers have less specific (/30) backup static routes pointing to their local servers

Limitations:

  • The server "self-monitor" itself, if it fails in a way where BGP is up, but DNS is unreachable from outside to the VIP (eg. iptables) this will cause an outage
  • By the nature of Anycast, Icinga will only check the health of the VIP closer to it
    • This could be worked around by checking the anycasted service health from various vantage points - task T311618
    • health checks to the servers' real IP still works

Deploying a new service

  1. Assign an IP in DNS, from the 10.3.0.0/24 range - (eg. Gerrit CR 524045)
  2. Configure the server side (eg. Gerrit CR 524037)
    1. Add include ::profile::bird::anycast where you see fit (usually to the service's role)
    2. Configure the VIP and its required attributes (usually hieradata/role/common/
      profile::bird::advertise_vips:
  <vip_fqdn>:  # used as identifier
    address: 10.3.x.x # VIP to advertise
    check_cmd: '/bin/true' # Any command to check the health of the service
    service_type: foobar # Can be any string, if underling applications need to distinguish VIPs
    check_fail: 2 # (Optional, default = 1) number of tries before the service is considered down
      Notes:
      • check_cmd is ran once per second from user "bird"
      • If any complex commands create a small bash script or use check_cmd: "/bin/sh -c '<commands>'"
      • anycast-healthchecker use the return code of the heath-check script, 0 = good, everything else is considered as a failure
      • IPv6 is supported but not enabled by default. See the section below on how to enable it.
  3. Configure the router side:
    1. set protocols bgp group Anycast4 neighbor <server_IP>
  4. Add monitoring to the VIP, similar to any Icinga checks, but in modules/profile/manifests/bird/anycast_monitoring.pp
  5. (Optional) if deploying a new type of service, ask Netops to add a backup static route

Other relevant configurations

Hiera keys:

# service the anycast-healthchecker binds to.
# bird is automatically bounded to anycast-healthchecker
profile::bird::bind_anycast_services: 
  - haproxy.service

# Router IPs with which Birds establish BGP sessions
# Usually set per site
profile::bird::neighbors_list:
  - routerIP
  - other_router_IP

# Fast failure detection between router and server (Optional, true by default)
profile::bird::bfd: true

# Usually set per service (role)
# But can be set for a specific host as well, for example to specifically remove the VIP from a host to be decommissioned.
profile::bird::advertise_vips:
  <vip_fqdn>: # Used as identifier
    address: 10.3.x.x # VIP to advertise (required)
    check_cmd: '/bin/true' # Any command to check the healh of the service, ran as user "bird" once per second (required)
    service_type: foobar # Can be any string, if underling applications need to distinguish VIPs (required)
    ensure: present # Set to absent to cleanly remove the check (optional, present by default)

# IPv6 support (experimental!)
# IPv6 is not enabled by default and needs to be explicitly enabled as the use case is limited and current deployments use IPv4
# To enable IPv6 support, you need to set do_ipv6 to true and then set the relevant IPv6 settings (address_ipv6 and check_cmd_ipv6)
profile::bird::do_ipv6: true
profile::bird::advertise_vips:
  <vip_fqdn>: # Used as identifier
    address: 10.3.x.x # VIP to advertise (required)
    check_cmd: '/bin/true' # Any command to check the healh of the service, ran as user "bird" once per second (required)
    service_type: foobar # Can be any string, if underling applications need to distinguish VIPs (required)
    ensure: present # Set to absent to cleanly remove the check (optional, present by default)
    address_ipv6: 2620:0:860 # /128 IPv6 VIP to advertise (required if do_ipv6 is set to true)
    check_cmd_ipv6: '/bin/true' # command to check the health of the service, for IPv6

Routing

Configuration

# show protocols bgp group Anycast4 
type external;
/* T209989 */
multihop {
    ttl 193;
}
local-address 208.80.153.193; # Router's loopback
damping;
import anycast_import;  # See below
family inet {
    unicast {
        prefix-limit {
            maximum 50; # Take the session down if more than 50 prefixes
            teardown;  # learned from the servers (eg. missconfiguration)
        }
    }
}
export NONE;
peer-as 64605;  # Server's ASN
bfd-liveness-detection {
    minimum-interval 300; # Take the session down after 3*300ms failures
}
multipath;  # Enable load balancing (remove for active/passive)
neighbor 208.80.153.111;  # Servers IPs
neighbor 208.80.153.77;   


# show policy-options policy-statement anycast_import 
term anycast4 {
    from {
        prefix-list-filter anycast-internal4 longer; # Only accept prefixes in the defined range
    }
    then {
        damping default;
        accept;
    }
}
then reject;

# show policy-options prefix-list anycast-internal4      
10.3.0.0/24;

# show routing-options static route 10.3.0.0/30 
next-hop 208.80.153.111;
readvertise;
no-resolve;

Routing to specific VIPs

Here both next hops (servers) are load balanced, as they are under the same *[BGP] block.

> show route 10.3.0.1
10.3.0.1/32        *[BGP/170] 1w4d 08:54:21, localpref 100, from 208.80.153.77
                      AS path: 64605 I, validation-state: unverified
                      to 208.80.153.77 via ae3.2003
                    > to 208.80.153.111 via ae4.2004

MTR can also be used for less granularity (site). Eg:

bast5001:~$ mtr 10.3.0.1 --report
Start: Fri Apr  5 16:48:21 2019
HOST: bast5001                    Loss%   Snt   Last   Avg  Best  Wrst StDev
  1.|-- ae1-510.cr2-eqsin.wikimed  0.0%    10    0.3   0.7   0.2   4.4   1.1
  2.|-- ae0.cr1-eqsin.wikimedia.o  0.0%    10    0.2   1.0   0.2   7.8   2.3
  3.|-- xe-5-1-2.cr1-codfw.wikime  0.0%    10  195.1 195.3 195.1 196.5   0.3
  4.|-- recdns.anycast.wmnet       0.0%    10  195.1 195.1 195.1 195.1   0.0

Monitoring

anycast_healthchecker logs can be viewed at /var/log/anycast-healthchecker/anycast-healthchecker.log

Bird's health can be monitored on its Grafana dashboard.

Syslog

For more context, see the main Gerrit change request (524037)

A limitation of a non-anycast setup: Some appliances such as network devices or PDUs can only send syslog to udp endpoints and not through the regular pipeline. The previous setup relied on 2 endpoints: syslog.codfw.wmnet and syslog.eqiad.wmnet, both CNAMEs. Some of those devices resolve the configured endpoint FQDN when the configuration is applied. This causes two issues:

  1. Changing the CNAME doesn't make the device send logs to the new endpoint
  2. Using DNS round robin is not possible

This left us with two options:

  1. Configure the devices with only 1 endpoint (eg. the geographically closer): which mean SPOF
  2. Configure the devices with both endpoints (not always supported): duplicated data in syslog

Using anycast, if an endpoint goes down, logs will automatically be routed to any other one.

Configuration

The VIP 10.3.0.4 (syslog.anycast.wmnet) is advertised by role syslog::centralserver (centrallog2002.codfw.wmnet and centrallog1001.eqiad.wmnet in Dec 2021)

Anycast healthchecker looks if there is a process listening on port udp/10514.

The rsyslog daemon on the centrallog hosts binds port 10514 UDP with a plaintext syslog listener, and forwards any logs recieved on this port along to the kafka logging/logstash pipeline using a config called "netdev_kafka_relay" or "netdev-kafka-relay"

Netconsole

The syslog hosts also run a Linux netconsole server to receive UDP logs from kernel urgent messages, the syslog anycast IP address is used by default. See Netconsole for more information.

Pooling/Depooling

To temporarily depool a service, disable Puppet, then stop bird.service.

To depool a server long-term, either:

  • Deactivate the neighbor IP on the router side
  • (Cleaner) Add a specific profile::bird::advertise_vips with the same identifier to the server, and check_cmd: /bin/false or ensure: absent

Upgrading Bird

See The bird 2 upgrade task for https://phabricator.wikimedia.org/T310574 for possible pitfalls.

Notably make sure the anycast-healthchecker and prometheus-bird-exporter packages and tools are compatible.

anycast-healthchecker logging

In the default configuration, anycast-healthchecker sets logging level to info and saves eight backups of logs to disk, taking care of log rotation itself. Since this may not be desired for hosts where anycast-hc is already functioning, you can decrease the verbosity and change the number of backups it maintains by using this Hiera configuration: 

 profile::bird::anycasthc_logging:
   level: 'critical'
   num_backups: 2

You can choose from the following logging levels: 'debug', 'info', 'warning', 'error', 'critical'.

Unicast announcements

You can announce unicast addresses with bird as well, using the existing setup, even though we use the word anycast because of the package name itself. https://phabricator.wikimedia.org/T348041 documents unicast announcements for the ns0 and ns1 nameservers.

Please make sure to announce the unicast address from its respective site and not from all sites by having an override in hieradata/role/$site/dnsbox.yaml. See commit https://gerrit.wikimedia.org/r/c/operations/puppet/+/965187/ for an example.

Troubleshooting

Bird daemon not running

This will trigger an automatic de-pool of the faulty server. Unless this is happening to multiple servers it is not an emergency.

Open a netops task and investigate like any daemon issue.

Has it been shutdown on purpose? Did any change have been made recently?

What do the log says (grep for "bird" in /var/log/syslog)?

Can it be restarted with service bird start ?

Anycast healthchecker not running

Same as above.

Logs are in /var/log/anycast-healthchecker/anycast-healthchecker.log

Process is sudo service anycast-healthchecker status

VIP not being announced by BGP

The VIP announcement status is controlled by the anycast-healthchecker mechanism.

In particular, if the command configured in check_cmd fails, the VIP will be removed from the BGP announcement.

How this works is like this:

  • the anycast-healthchecker service reads the configuration for VIPs from /etc/anycast-healthchecker.d/*.conf files
  • for each file, it runs the check_cmd command.
  • the service will add/remove the VIP from the /etc/bird/anycast-prefixes.conf file dynamically based on the result of the check.
Retrieved from "https://wikitech.wikimedia.org/w/index.php?title=Anycast&oldid=2119730"