Live Migration of Rescued Instances¶
Add support to live migrate rescued instances. Currently instances in rescued state cannot be live migrated. This might be an issue during upgrades as all the active instances in the host needs to be migrated using live-migrate.
Currently if an operator wants to upgrade a host with rescued instance, it needs to be unrescued and live migrated before proceeding with maintenance operation.
During rescue operation, libvirt driver creates set of files in the instance directory. These files are used when unrescuing the instance to revert to the original disk configuration. During live migration of rescued instance, libvirt only moves files to destination host that it knows of, leaving behind the files created by libvirt driver in source host. This causes unrescue to fail on destination node.
This implementation will allow live migration of rescued instances for libvirt driver and add driver capabilities flag to check if the driver supports live migration of rescued instances.
Operators would like to live migrate all running instances including VM’s that are currently in rescued state before performing any maintenance activities on the host and ensuring 100% uptime for all VM’s in running state.
Operators prefer to live migrate rescued instance without the hassle of unrescuing them before live migration. This creates better user experience to the operator and reduces any service interruptions to the end user.
A new API microversion will be introduced to allow live migration of instances in rescued state. In the new microversion, minimum nova-compute service version in the deployment is checked for the support of live migration of rescued instances else the requests fails at the API level. This is useful during upgrade when some compute nodes may not have been upgraded to the version that supports the migration of rescued instances.
A new driver capability flag called supports_live_migrate_rescued will be added to all drivers and set to True or False depending on their ability to live migrate rescued instance. This spec will only enable this for libvirt driver. All other drivers will fail to live-migrate if you attempt to live-migrate an instance in the rescued state. Driver’s support to live migrate rescued instance is checked at pre-live migration phase which is an async operation. Any failures due to driver’s capability to support live migration of rescued instance, will be updated in instance-actions with the error message, just as we do for any other pre-live migration errors.
Presently in live migration, files that are created by libvirt driver while rescuing instance are not copied to the target compute node. The proposal is to copy unrescue.xml from source to destination host, download kernel.rescue and ramdisk.rescue if they exist on image service else fallback to copy those files from source host. These operations are carried out during pre-live migration phase.
One alternative is, leaving it up to operators to ask the user to unrescue their instance so it can be migrated.
Another option is to write the original instance state to the database during the rescue operation so that it can be retrieved by the target compute node during unrescue. We are considering this idea separately to this spec around live-migration.
Data model impact¶
REST API impact¶
A new microversion is required because os-migrateLive action API call will return a response code of 202 instead of the current 409 response.
When not all compute nodes are upgraded to minimum compute version that supports this functionality 409 response is returned.
Other end user impact¶
Bump python-novaclient API version.
Other deployer impact¶
This features is available only when all compute nodes are upgraded.
- Primary assignee:
- Other contributors:
Implement driver capability check to verify if the driver supports live migration of rescued instances and copy the rescued instance files created by libvirt driver to the destination host prior to migration
Change compute API live migration methods to allow migration of rescued instances and check for minimum compute version across deployments.
Bump python-novaclient API version.
Unit tests will be added as required.
Add tempest tests to verify the use of live migration of an instance in a rescued state and subsequent unrescuing of the instance.
Need to document API changes in api-ref:
Compute API extensions documentation http://developer.openstack.org/api-ref-compute-v2.1.html
Reproposed and Updated