Jump to content


From Wikitech
Wikimedia infrastructure

Data centres and PoPs


HTTP Caching




Wikimedia uses LVS for balancing traffic over multiple servers, see also load balancing architecture


File:Esams LVS.png
Diagram of LVS setup in Esams, anno 2010

We use LVS-DR, or Direct Routing. This means that only forward (incoming) traffic is balanced by the load balancer, and return traffic does not even go through the load balancer. Essentially, the LVS balancer receives traffic for a given service IP and port, selects one out of multiple "real servers", and then forwards the packet to that real server with only a modified destination MAC address. The destination servers also listen to and accept traffic for the service IP, but don't advertise it over ARP. Return traffic is simply sent directly to the gateway/router.

The LVS balancer and the real servers need to be in the same subnet for this to work.

The real servers are monitored by a Python program called Pybal. It does certain kinds of health checks to determine which servers can be used, and pools and depools them accordingly. You can follow what Pybal is doing in log file /var/log/pybal.log.

PyBal also has an integrated BGP module that Mark has written (Twisted BGP, available in the MediaWiki SVN repository). This is used as a failover/high availability protocol between the LVS balancers (PyBal) and the routers. PyBal announces the LVS service IPs to the router(s) to indicate that it is alive and can serve traffic. This also removes the need to manually configure the service IPs on the active balancers. All LVS servers are now using this setup.

HOWTO (Make sure you know whether you are using Etcd or not!)

Etcd as a backend to Pybal (Codfw only)

In order to manage Pybal pools in Etcd use Conftool and confctl.

High level overview:

  • Define node and services in conftool-data/ in ops-puppet
  • puppet-merge and conftool-merge your change
  • Nodes will usually inherit a default pool/weight value based on their service default
  • To modify the state of the Node per service in Etcd use confctl
  • Confd is running on the LVS hosts and using HTTP Long Polling to watch for changes in the service definition in Etcd. Any change should be picked up by Confd and used to modify the on disk node pool defined in /etc/pybal/pools/.
  • This change is polled periodically in Pybal (currently every 60s)

Pool or depool hosts (for non-Etcd managed pools)

Edit the files in /srv/pybal-config/pybal/$colo on config-master.$colo and wait a minute - PyBal will fetch the file over HTTP. Please don't forget to commit your changes locally.

If you set a host do disabled, PyBal will continue to monitor it but just not pool it:

{ 'host': 'knsq1.esams.wikimedia.org', 'weight': 10, 'enabled': False }

If you comment the line, PyBal will forget about it completely.

In emergency cases, you can do this manually using ipvsadm, if PyBal for some reason is not working for example.

ipvsadm -d -t VIP:PORT -r REALSERVER

Such as:

ipvsadm -d -t -r knsq1.esams.wikimedia.org

Note that PyBal won't know about this, so make sure you bring the situation back in sync.

Example request for checking the current status for restbase in eqiad via http:

curl http://config-master.eqiad.wmnet/pybal/eqiad/restbase

See which LVS balancer is active for a given service

If you have ssh access to the host in question, sshing to the IP address will land you in a shell on whichever system is active.

 $ ssh root@ms-fe.pmtpa.wmnet

If you don't want to connect (or can't connect) to the system, ask the directly attached routers. You can request the route for a given service IP. E.g. on Foundry:

csw1-esams#show ip route
Type Codes - B:BGP D:Connected I:ISIS S:Static R:RIP O:OSPF; Cost - Dist/Metric
Uptime - Days:Hours:Minutes:Seconds 
        Destination        Gateway         Port        Cost     Type Uptime
1  ve 1        20/1     B    10:14:28:44

So (amslvs2] is active for Upload LVS service IP

On Juniper:

csw2-esams> show route 

inet.0: 38 destinations, 41 routes (38 active, 0 holddown, 0 hidden)
+ = Active Route, - = Last Active, * = Both  *[BGP/170] 19:38:18, localpref 100, from
                      AS path: 64600 I
                    > to via vlan.100
                    [BGP/170] 1w3d 14:24:52, MED 10, localpref 100
                      AS path: 64600 I
                    > to via vlan.100

So (*) is active for Text LVS service IP

To see all LVS servers configured for a service

To see which servers are configured for a service, but not which server is currently active, look in the puppet configs.

  • configuration is stored in puppet/manifests/lvs.pp
  • in the class lvs::configuration
  • in the $lvs_services variable,
  • look for your service (eg 'swift' or 'upload')
  • look for the class, which will be something like low-traffic
  • back up in lvs.pp to the section defining the $lvs_class_hosts variable.
  • look for your class (eg low-traffic)
  • you should see sections for production and labs, with variables for each data center listing the lvs servers responsible.

example showing that lvs3 and 4 are responsible for swift (command output excerpted for clarity):

 ben@fenari:~/pp/manifests$ grep -A2 "swift" lvs.pp 
               "swift" => {
                       'description' => "Swift object store for thumbnails",
                       'class' => "low-traffic",
 ben@fenari:~/pp/manifests$ grep -A5 "low-traffic" lvs.pp  | head -n 5
               'low-traffic' => $::realm ? {
                       'production' => $::site ? {
                               'pmtpa' => [ "lvs3", "lvs4" ],
                               'eqiad' => [ "lvs1003", "lvs1006" ],
                               'esams' => [ ],

Deploy a change to an existing service


  • you have already made the change in puppet and puppet-merged it on the puppet master (palladium).
  • you have tested the change on the backend real servers directly (eg if you were changing a health check URL you have already queried the backend servers for that URL successfully).

Deploy steps:

  • find out which LVS servers host your service (see above). For this example, I'll use lvs 3 and 4.
  • find out which LVS server is active (see above). For this example, I'll assume lvs4.
  • log into the inactive host twice.
  • in one session, tail the pybal log looking for one (or more) of your backend servers. eg tail -f /var/log/pybal.log | grep ms-fe
    • you should see one line per server every 10 seconds or so. Look for "enabled/up/pooled". eg:
 2012-03-13 19:12:20.003332 [ProxyFetch] ms-fe2.pmtpa.wmnet (enabled/up/pooled): Fetch successful, 0.006 s
  • in the other session, run puppet and verify your change exists on the local filesystem
  • double check /home/wikipedia/conf/pybal/
  • get a list of all IP addresses served by this LVS server - you're going to check that they all exist after your change
    • run ip addr and save the output for later
  • restart pybal:
    • /etc/init.d/pybal stop
    • If pybal doesn't die correctly, which shouldn't happen anymore, kill it with SIGKILL: pkill -9 pybal
    • /etc/init.d/pybal start
  • check that all the expected IP addresses exist
    • run 'ip addr' and compare against the list you collected before making your change
  • in the log you're tailing, you should see a few messages like:
 2012-03-13 19:21:26.015393 New enabled server ms-fe1.pmtpa.wmnet, weight 40
 2012-03-13 19:21:26.015611 New enabled server ms-fe2.pmtpa.wmnet, weight 40
 2012-03-13 19:21:26.015666 ['-a -t -r ms-fe2.pmtpa.wmnet -w 40', '-a -t -r ms-fe1.pmtpa.wmnet -w 40']
 2012-03-13 19:21:26.074253 [IdleConnection] ms-fe1.pmtpa.wmnet (enabled/up/pooled): Connection established.
 2012-03-13 19:21:26.075302 [IdleConnection] ms-fe2.pmtpa.wmnet (enabled/up/pooled): Connection established.
 2012-03-13 19:21:36.022517 [ProxyFetch] ms-fe1.pmtpa.wmnet (enabled/up/pooled): Fetch successful, 0.007 s
 2012-03-13 19:21:36.023442 [ProxyFetch] ms-fe2.pmtpa.wmnet (enabled/up/pooled): Fetch successful, 0.008 s
 2012-03-13 19:21:46.030482 [ProxyFetch] ms-fe1.pmtpa.wmnet (enabled/up/pooled): Fetch successful, 0.008 s
  • Look for "Fetch successful". This is the line that means your change is successful. Seeing 'enabled/up/pooled' is insufficient.
    • example of a failed change (note that it still says enabled/up/pooled for a few lines - this does not mean that it's ok! look for the Fetch line):
 2012-03-13 19:27:23.626787 [IdleConnection] ms-fe2.pmtpa.wmnet (enabled/up/pooled): Connection established.
 2012-03-13 19:27:23.632928 [IdleConnection] ms-fe1.pmtpa.wmnet (enabled/up/pooled): Connection established.
 2012-03-13 19:27:33.555879 [ProxyFetch] ms-fe2.pmtpa.wmnet (enabled/up/pooled): Fetch failed, 0.005 s
 2012-03-13 19:27:33.555917 Monitoring instance ProxyFetch reports servers ms-fe2.pmtpa.wmnet (enabled/up/pooled) down: 404 Not Found
 2012-03-13 19:27:33.556022 ['-d -t -r ms-fe2.pmtpa.wmnet']
 2012-03-13 19:27:33.562458 [ProxyFetch] ms-fe1.pmtpa.wmnet (enabled/up/pooled): Fetch failed, 0.012 s
 2012-03-13 19:27:33.562533 Monitoring instance ProxyFetch reports servers ms-fe1.pmtpa.wmnet (enabled/up/pooled) down: 404 Not Found
 2012-03-13 19:27:33.562589 Could not depool server ms-fe1.pmtpa.wmnet because of too many down!
 2012-03-13 19:27:43.561745 [ProxyFetch] ms-fe2.pmtpa.wmnet (enabled/partially up/not pooled): Fetch failed, 0.002 s
 2012-03-13 19:27:43.565608 [ProxyFetch] ms-fe1.pmtpa.wmnet (enabled/partially up/pooled): Fetch failed, 0.003 s
  • if your change was successful, repeat the procedure on the active host.
    • when you stop pybal on the active host, traffic will immediately fail over to the standby host.
    • when you restart pybal on the formerly active host, traffic will immediately fail back (the LVS pairs are configured with a default and a standby so traffic always flows to the default if it's up).

Add a new load balanced service

First choose whether it's a 'high-traffic' (aka public-facing) or 'low-traffic' (aka internal post-cache)

DNS changes

  • allocate an IP address per colo to serve your content
  • internal addresses should have names *.svc.$colo.wmnet:
    • codfw should be in the range
    • eqiad should be in the range
  • external addresses:
    • These need to be allocated from the (small!) public IP address pool, and may need specific configuration on the routers. Talk to the network admins first (Mark/Faidon).

Puppet Changes

  • in modules/lvs/manifests/configuration.pp
    • add a stanza to $lvs_service_ips defining the IP addresses mimicking an existing entry
    • add a stanza to $lvs_services mimicking an existing entry
      • use your choice of high or low volume traffic here
  • in manifests/role/lvs.pp
    • add the IP addresses to $lvs_balancer_ips in the section for the lvs servers that will serve your service.

noc changes

  • create a file in /home/wikipedia/conf/pybal/$colo for your service listing which hosts are available for the backend
  • see PyBal for more detail on how this part works

Deploy your changes

  • see instructions above (Deploy a change to an existing service)

LVS installation

LVS now uses Puppet and automatic BGP failover. Puppet arranges the service IP configuration, and installation of packages. To configure the service IPs that an LVS balancer should serve (both primary and backup!), set the $lvs_balancer_ips variable:

node /amslvs[1-4]\.esams\.wikimedia\.org/ {
        $cluster = "misc_esams"

        $lvs_balancer_ips = [ "", "", "", "" ]

        include base,

In this setup, all 4 hosts amslvs1-amslvs4 are configured to accept all service IPs, although in practice every service IP is only ever serviced by one out of two hosts due to the router configuration.

Puppet uses the (now misleadingly named) wikimedia-lvs-realserver package to bind these IPs to the loopback (!) interface. This is to make sure that a server answers on these IPs, but does not announce them via ARP - we'll use BGP for that.

LVS service configuration

In file lvs.pp the services themselves are configured, from which the PyBal configuration file /etc/pybal/pybal.conf is generated by Puppet.

Most configuration is in a large associative hash, $lvs_services. Each key in this hash is the name of one LVS service, and points to hash of PyBal configuration variables:

Textual description of the LVS service.
The class the LVS service belongs too; i.e. on which LVS balancers it is active (see below).
A hash of service IP address for the service. All IP addresses are aliases, and are translated to separate LVS services in PyBal.conf, but with identical configuration.

The other configuration variables are described in the PyBal article.

Global PyBal configuration options can be specified in the $pybal hash.


To determine which LVS services are active on which hosts, the $lvs_class_hosts determines for each class, which hosts should have the services for that class. This is used by the pybal.conf template to generate the LVS services. The following classes are used, to distribute traffic over the LVS balancer hosts:

  • high-traffic1 (text, bits)
  • high-traffic2 (text, upload)
  • https (HTTPS services corresponding to the 'high-traffic' HTTP services; should be active on all hosts that carry either class)
  • specials (special LVS services, especially those that do not have BGP enabled)
  • low-traffic (internal load balancing, e.g. from the Squids to the Apaches)

BGP failover and load sharing

Previously, the LVS balancer that had a certain service IP bound to its eth0 interface was active for that IP. To do failovers, the IP had to be moved manually.

In the new setup, multiple servers announce the service IP(s) via BGP to the router(s), which then pick which server(s) to use based on BGP routing policy.

PyBal BGP configuration

In the global section, the following BGP related settings typically exist:

bgp = yes

Enables bgp globally, but can be overridden per service.

bgp-local-asn =  64600

The ASN to use while communicating to the routers. All prefixes will get this ASN as AS path.

bgp-peer-address =

The IP of the router this PyBal instance speaks BGP to.

#bgp-as-path = 64600 64601

An optional modified AS path. Can be used e.g. to make the AS path longer and thus less attractive (on a backup balancer).

Example BGP configuration for Juniper (cr1-esams)

mark@cr1-esams> show configuration protocols bgp group PyBal 
type external;
multihop {
    ttl 2;
hold-time 30;
import LVS_import;
family inet {
    unicast {
        prefix-limit {
            maximum 50;
family inet6 {
    unicast {
        prefix-limit {
            maximum 50;
export NONE;
peer-as 64600;

mark@cr1-esams> show configuration policy-options prefix-list LVS-service-ips   ;;

mark@cr1-esams> show configuration policy-options prefix-list LVS-service-ips6   

mark@cr1-esams> show configuration routing-options aggregate                               

mark@cr1-esams> show configuration routing-options rib inet6.0 aggregate 
route 2620:0:862:ed1a::/64;

mark@cr1-esams> show configuration policy-options policy-statement ospf_export 
term 1 {
    from protocol direct;
    then accept;
term statics {
    from protocol [ static aggregate ];
    then accept;
then reject;

The LVS_import policy adds metric 10 to the "routes" (service IPs) received from the secondary (backup) LVS balancers. This means that the router will regard them as less preferred.

The individual /32 and /128 service IPs are announced by PyBal and exchanged between the routers using IBGP. Aggregates for the service IP ranges are generated by the core routers and redistributed into OSPF as well.

SSH checking

As the Apache cluster is often suffering from broken disks which break SSH but keep Apache up, I have implemented a RunCommand monitor in PyBal which can periodically run an arbitrary command, and check the server's health by the return code. If the command does not return within a certain timeout, the server is marked down as well.

The RunCommand configuration is in /etc/pybal/pybal.conf:

runcommand.command = /bin/sh
runcommand.arguments = [ '/etc/pybal/runcommand/check-apache', server.host ]
runcommand.interval = 60
runcommand.timeout = 10
The path to the command which is being run. Since we are using a shell script and PyBal does not invoke a shell by itself, we have to do that explicitly.
A (Python) list of command arguments. This list can refer to the monitor's server object, as shown here.
How often to run the check (seconds).
The command timeout; after this amount of seconds the entire process group of the command will be KILLed, and the server is marked down.

Currently we're using the following RunCommand script, in /etc/pybal/runcommand/check-apache:


set -e

SSH_OPTIONS="-o PasswordAuthentication=no -o StrictHostKeyChecking=no -o ConnectTimeout=8"

# Open an SSH connection to the real-server. The command is overridden by the authorized_keys file.
ssh -i /root/.ssh/pybal-check $SSH_OPTIONS $SSH_USER@$HOST true

exit 0

The limited ssh accounts on the application servers are managed by the wikimedia-task-appserver package.


Nagios monitoring of LVS services is managed by Puppet as well, at the bottom of the lvs.pp file. For example:

monitor_service_lvs_http { "wikimedia-lb.pmtpa.wikimedia.org":
    ip_address => "",
    check_command => "check_http_lvs!meta.wikimedia.org!/wiki/Main_Page"
monitor_service_lvs_https { "wikimedia-lb.pmtpa.wikimedia.org":
    ip_address => "",
    check_command => "check_https_url!meta.wikimedia.org!/wiki/Main_Page"

Diagnosing problems

Run ipvsadm -l on the director. Healthy output looks like this:

IP Virtual Server version 1.2.1 (size=4096)
Prot LocalAddress:Port Scheduler Flags
  -> RemoteAddress:Port           Forward Weight ActiveConn InActConn
TCP  upload.pmtpa.wikimedia.org:h wlc
  -> sq10.pmtpa.wmnet:http        Route   10     5202       5295
  -> sq1.pmtpa.wmnet:http         Route   10     8183       12213
  -> sq4.pmtpa.wmnet:http         Route   10     7824       13360
  -> sq5.pmtpa.wmnet:http         Route   10     7843       12936
  -> sq6.pmtpa.wmnet:http         Route   10     7930       12769
  -> sq8.pmtpa.wmnet:http         Route   10     7955       11010
  -> sq2.pmtpa.wmnet:http         Route   10     7987       13190
  -> sq7.pmtpa.wmnet:http         Route   10     8003       7953

All the servers are getting a decent amount of traffic, there's just normal variation.

If a realserver is refusing connections or doesn't have the VIP configured, it will look like this:

IP Virtual Server version 1.2.1 (size=4096)
Prot LocalAddress:Port Scheduler Flags
  -> RemoteAddress:Port           Forward Weight ActiveConn InActConn
TCP  upload.pmtpa.wikimedia.org:h wlc
  -> sq10.pmtpa.wmnet:http        Route   10     2          151577
  -> sq1.pmtpa.wmnet:http         Route   10     2497       1014
  -> sq4.pmtpa.wmnet:http         Route   10     2459       1047
  -> sq5.pmtpa.wmnet:http         Route   10     2389       1048
  -> sq6.pmtpa.wmnet:http         Route   10     2429       1123
  -> sq8.pmtpa.wmnet:http         Route   10     2416       1024
  -> sq2.pmtpa.wmnet:http         Route   10     2389       970
  -> sq7.pmtpa.wmnet:http         Route   10     2457       1008

Active connections for the problem server are depressed, inactive connections normal or above normal. This problem must be fixed immediately, because in wlc mode, LVS load balances based on the ActiveConn column, meaning that servers that are down get all the traffic.

Incorrectly bound interfaces

Don't ever bind IP addresses directly to lo in /etc/network/interfaces. If you do, stuff breaks. (This applies not just to LVS servers but any real server as well. Anything with the wikimedia-lvs-realserver package will break if you bind addresses manually.)

When it's broken, it looks like this. Notice that all the balanced IP addresses are tagged lo:LVS except 13 is broken and causes the ifup script that reloads the IPs to be broken.

    inet scope host lo
    inet scope global lo
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet6 ::1/128 scope host 

The solution here is to delete the broken interface (ip addr del dev lo), then run dpkg-reconfigure wikimedia-lvs-realserver. This triggers the scripts that will re-add all the IP addresses.

Happy ip addr output looks like this:

root@lvs4:/etc/network# ip addr
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 16436 qdisc noqueue state UNKNOWN 
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet scope host lo
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet scope global lo:LVS
    inet6 ::1/128 scope host 
       valid_lft forever preferred_lft forever
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP qlen 1000