Portal:Cloud VPS/Bare Metal

From Wikitech

The concept of Cloud VPS 'bare metal' has taken on a life of its own and caused confusion in the tradition of Labs labs labs. This page attempts to explain some of the history behind it, and also state the current position of the Cloud Services team.

Current status

"Bare metal" instances are currently not provided by Wikimedia Cloud VPS.

History

The Cloud Services team was interested in Openstack Ironic and once had a quarterly goal for investigating the feasibility of using Openstack to manage and allocate physical hardware for tenant use. Ironic depends on Neutron which we figured out quickly was enough work on its own. We shifted focus to determine how maintainable it would be to hack our own solution for the same basic purpose. To that end we acquired a PoC server that was allocated into the instance VLAN. We hacked in DNS via LUA, and created DHCP reservations.

At the end of the quarterly goal we had a pilot host that somewhat functioned as expected. We did not sort out console access, security issues, support or really anything formal for general use. The long tail of this is essentially recreating Ironic on our own. We agreed as a team (coren/andrew/yuvi/chase) that this is not a supportable model at our level of staffing and to wait for Ironic and reassess. We have only the initial allocations from PoC which we kept and let subbu use with the understanding that it was not supported and would eventually be phased out.

We have talked about a possible specific exception if a compelling case is made but so far this has not happened.

Capabilities & Limitations of the Proof of Concept

  • Physical machine allocated for tenant use within the instance VLAN
  • Uses the Cloud VPS puppetmaster or the puppetmaster of the appropriate tenant
  • Cannot have NFS (wasn't part of the PoC)
  • Has no association with security groups as it is our own hackery
  • Has DNS/DHCP hacked in
  • Is imaged from production carbon installer
  • Is not tied in with any of the Openstack ecosystem (as such things like floating IPs do not work)
  • Is something the Cloud Services team agreed is not a supportable model currently

Things that could be confused with 'bare metal in Cloud VPS'

  • Servers in the labs-support VLAN for production services (not part of the Openstack ecosystem) in support of Cloud VPS instances (production hardware supporting Cloud VPS instances - labsdb, cloudmetrics, etc)
  • Cloud VPS/Openstack hardware that runs the components to keep Cloud VPS running like nova or designate. (production hardware that powers Cloud VPS)
  • Servers in standard production VLANs providing services to Cloud VPS instances such as ldap (production hardware or ganeti VM's that support core Cloud VPS functionality)
  • Servers acting as OpenStack Compute (nova) nodes which as not in the normal scheduler pool and instead hosting instances specifically placed there by OpenStack administrators (such as some of the clouddb-services project instances).