Labs infrastructure

From Wikitech

How Labs Works

Labs is a virtualization cloud that uses OpenStack Compute. Base images are managed with Glance and authentication uses LDAP-backed Keystone.

Labs currently runs in a single datacenter in Ashburn, Virginia. In the future it will span two or more datacenters, with a slightly different configuration in each.

For troubleshooting immediate labs issues, visit the labs_troubleshooting page.

Slides from a brief presentation about Wikimedia Labs architecture

Labs Eqiad (Ashburn, VA)

File:Labs cluster.png
The servers that make Wikimedia Labs

Wikitech/OpenStackManager

In addition to hosting the WMF's technical documentation, the wikitech web site also runs the OpenStackManager MediaWiki extension that provides a graphical interface for labs. Wikitech runs on a server internally named Silver.

An alternative, partially-functional labs GUI can be accessed at https://horizon.wikimedia.org/. It runs the openstack-dashboard project and provides nova API access to project admins.

Controller

The labs controller box (currently named 'labcontrol1001') runs the Glance and Keystone services, as well as a few nova services (conductor and scheduler.) Labcontrol1001 also runs a public DNS server (aka labs-ns0) which provides name services for the .wmflabs.org domain.

A second server, labcontrol1002, serves as a hot spare for labcontrol1001.

Another duplicate server, labcontrol2001, runs in codfw and contains a duplicate config. It is largely vestigial, but does provide backup DNS service via the labs-ns1 service name.

Network

The network node ('labnet1001') hosts the nova-network service. We currently run a single labs-wide network that supports all lab nodes and projects. In the future we hope to use Openstack Neutron for our network setup, but it doesn't support our use-case; to use neutron we'll need to switch to one network per project.

Labnet1001 also hosts the nova-api service.

Soon an additional server, labnet1002, will provide either redundant network service or as a hot spare. To be determined.

Virtualization

File:Labs cluster instance POV.png
The servers that a labs instance talks to

There are currently nine virtualization nodes in labs, named labvirt1001-1009, all running in eqiad:

  • 1001-1009 are high-powered multi-CPU HP servers; each of them hosts dozens of virtual machines.

Storage

Labs uses shared storage for several purposes:

  • Each member of a project has a project-wide shared home directory.
  • Each project has a public shared volume, generally mounted to /data/project

All of the above are hosted on an NFS server named labstore1001. There's a hot-swappable backup, labstore1002, which is generally turned off.

monitoring

Labmon1001 runs statsd and graphite. It monitors the state of labs instances and collects stats and sends alerts as needed.

ldap

Ldap is used for services throughout the WMF. The primary ldap server is Neptunium, running in eqiad. The secondary server is Nembus, running in codfw.

The LDAP server software is opendj. Each labs instance has an /etc/ldap.conf file (managed by puppet) that maintains info about the ldap servers.

dns

Labs DNS is handled by PowerDNS. Private DNS entries (e.g. foo.eqiad.wmflabs) are created via Designate Sink and stored in a PDNS server using a mysql backend. Public DNS is handled by a PDNS server backed by ldap, which is read and written by OpenStackManager.

File:Labs dns mixed.png
Old Labs DNS design using two different pdns implementations

Soon we will switch to Horizon for the public DNS web interface, at which point all DNS will be handled via Designate and single PDNS server.

File:Labs dns simplified.png
Future, simpler Labs DNS implementation using Horizon