Scheduling interaction for cells

https://blueprints.launchpad.net/nova/+spec/cells-scheduling-interaction

In order to schedule instance builds to compute hosts Nova and the scheduler will need to take into account that hosts are grouped into cells. It is not necessary that this is apparent when Nova is requesting a placement decision.

Problem description

In order to partition Nova into cells the scheduler will need to be involved earlier in the build process.

Use Cases

  • Operators want to partition their deployments into cells for scaling, failure domain, and buildout reasons. When partitioned, we need to have flexible api level scheduling that can make decisions on cells and hosts.

Proposed change

The scheduler will be queried by Nova at the api level so that it knows which cell to pass the build to. The instance table exists in a cell and not in the api database, so to create an instance we will first need to know which cell to create it in.

The scheduler will continue to return a (host, node) tuple and the calling service will look up the host in a mapping table to determine which cell it is in. This means the current select_destinations interface will not need to change. Querying the scheduler will take place after the API has returned a response so the most reasonable thing to do is pass the build request to a conductor operating outside of any cell. The conductor will call the scheduler and then create the instance in the cell and pass the build request to it.

While much of Nova is being split between being api-level or cell-level the scheduler remains outside of either distinction. It can be thought of as a separate service in the same way that Cinder/Neutron are. As a result the scheduler will require knowledge of all hosts in all cells. This is different from the cellsv1 architecture and may be a surprise for those familiar with that setup. A separate effort can address scalability for the scheduler and the potential for partitioning it along some boundary, which may be a cell.

Because the existing build_instances conductor method assumes that the instance already exists within a cell database and makes some assumptions about cleaning up resources on failure we will not complicate that method with an alternate path. Instead a new conductor method will be added which can take the task of querying the scheduler and then creating the instance and associated resources within the cell indicated by the scheduler.

The new boot workflow would look like the following:

  • nova-api creates and persists a BuildRequest object, not an Instance.

  • Cast to the api level conductor to execute the new method proposed here. The api level conductor is whatever is configured in DEFAULT/transport_url.

    • Conductor will call the scheduler once.

    • Conductor will create the instance in the proper cell

    • Conductor will cast to the proper nova-compute to continue the build process. This cast will be the same as what is currently done in the conductor build_instances method.

    • In the event of a reschedulable build failure nova-compute will cast to a cell conductor to execute the current build_instances method just as it’s currently done.

Rescheduling will still take place within a cell via the normal compute->conductor loop, using the conductors within the cell. Adding rescheduling at a cell level will be a later effort.

Information about cells will need to be fed into the scheduler in order for it to account for that during its reschedule/migration placement decisions, but that is outside the scope of this spec.

Alternatives

We could query the scheduler at two points like in cellsv1. This creates more deployment complexity and creates an unnecessary coupling between the architecture of Nova and the scheduler.

A reschedule could recreate a BuildRequest object, delete the instance and any associated resources in a cell, and then let the scheduler pick a new host in any cell. However there is a fair bit of complexity in doing that cleanup and resource tracking and I believe that this is an effort best left for a later time. It should also be noted that any time a build crosses a cell boundary there are potential races with deletion code so it should be done as little as possible.

Rather than requiring an api level set of conductor nodes nova-api could communicate with the conductors within a cell thus simplifying deployment. This is probably worth doing for the single cell case so another spec will propose doing this.

Data model impact

None

REST API impact

None

Security impact

None

Notifications impact

None

Other end user impact

None

Performance Impact

None

Other deployer impact

Nova-conductor will need to be deployed for use by nova-api.

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

alaski

Other contributors:

None

Work Items

  • Add a conductor method to call the scheduler, create an instance in the db of the cell scheduled to, then cast to the selected compute host to proceed with the build.

  • Update the compute api to not create the instance in the db during a build request, and change it to cast to the new scheduler method.

  • Ensure devstack is configured to that nova-api shares the cell level conductors. This makes the single cell setup as simple as possible. A later effort can investigate making this configurable in devstack for multiple cell setups.

Dependencies

None

Testing

Since this is designed to be an internal re-architecting of Nova with no user visible changes the current suite of Tempest or functional tests should suffice. At some point we will want to look at how to test multiple cells or potentially exposing the concept of a cell in the API and we will tackle testing requirements then.

Documentation Impact

Documentation will be written describing the flow of an instance build and how and where scheduling decisions are made.

References

https://etherpad.openstack.org/p/kilo-nova-cells https://etherpad.openstack.org/p/nova-cells-scheduling-requirements

History

Revisions

Release Name

Description

Liberty

Introduced

Mitaka

Re-proposed; partially implemented.

Newton

Re-proposed; partially implemented.

Ocata

Re-proposed.