volume-backed server rebuild¶
Currently, the compute API will fail if a user tries to rebuild a volume-backed server with a new image. This spec proposes to add support for rebuilding a volume-backed server with a new image.
Currently Nova rebuild (with a new image) only supports instances which are booted from images. The volume-backed instance cannot be rebuilt when a new image is supplied. Trying to rebuild a volume-backed instance will raise a HTTPBadRequest exception.
As a user, I would like to rebuild my volume-backed server with a new image.
As a nova developer, I would like to have feature parity in the compute API for volume-backed and image-backed servers.
First, change the existing API for rebuilding a volume-backed server. Then the API flow would be:
Has the new API microversion been requested?
Is the instance.host service version new enough to support volume-backed rebuild with a new image?
If these are true, proceed. If not, fail in the API with a 409 error.
Note that when rebuilding with a new image, the request will be run through the scheduler against the current host to be consistent with image-backed rebuild with a new image. See bug 1664931 for details.
Then the nova-compute will perform the following steps:
Create an empty (no connector) volume attachment for the volume and server. This ensures the volume remains
reservedthrough the next step.
Delete the existing volume attachment (the old one).
Call the new
Poll the volume status for completion (either success or failure).
Upon successful completion of the re-image operation, update the empty volume attchment in Cinder, and then do the attachment on the Nova host when spawning the (rebuilt) guest VM and “complete” the attachment which will make the volume
In this process, there are some conditions that we could hit:
If we failed to re-image the volume and the volume is in ‘error’ status then we should set the instance status as “error”. Since users can rebuild instances in error status, the user has a way to retry the rebuild once the cause of the cinder side failure is resolved. Note that nova-compute will not attempt to update the volume attachment records with the host connector again on the volume in error status.
If the cinder API itself returns a >=400 error, nothing changed about the root volume and in that case the migration status can be ‘failed’ but the instance status should go back to what it was (we can see how _error_out_instance_on_exception is used).
The main alternative is that nova would perform the rebuild like an initial boot from volume where nova-compute would create a new volume from the new image and then “swap” the root volume on the instance during rebuild.
There are issues with this, however, like what to do about the old volume:
Regarding ‘delete_on_termination’ flag in the BDM, delete_on_termination=True means: don’t delete the volume when we kill the instance. Rebuild means: re-initialize this instance in place. The rebuild flow would have to determine what to do if the old root volume BDM was marked with delete_on_termination=True - ignore that and preserve the old root volume or delete it.
We could pass a new flag to the rebuild API telling nova what to do about the old volume (delete it or not). If the flag is true to delete the old volume but the old volume has snapshots, Nova won’t be deleting the volume snapshots just to delete the volume during a rebuild.
But there are several issues with that as mentioned above like quota and the questions about what nova should do about the old volume, you can see more detailed information in References.
Data model impact¶
REST API impact¶
Change the rebuild request response code from 400 to 202 if the conditions described in the Proposed change section are met. The API microversion and compute RPC version will also be incremented to indicate the new support.
Other end user impact¶
The python-novaclient and python-openstackclient will be updated to support the new microversion.
The operation will take longer because of the orchestration involved and the work that needs to happen in Cinder.
Other deployer impact¶
If the cinder volume
reimage API operation fails and the volume goes to
error status, an admin will likely need to investigate and resolve the
issue in cinder and then reset the volume status to
The API microversion and compute RPC version will also be incremented to indicate the new support, therefore users will not be able to leverage the feature until the nova-compute service hosting a volume-backed instance is upgraded.
- Primary assignee:
Jie Li <email@example.com> (ramboman)
Change the existing rebuild API.
Create an empty attachment for the root volume so the volume remains in-use during rebuild (we do this today already).
Delete the old volume attachment.
Call the cinder API to re-image the volume.
Update and complete the volume attachment once re-imaged.
Adopt the new compute version.
Adopt the new microversion in python-novaclient.
Adopt the new microversion in python-openstackclient.
Change the nova API documents.
Depends on the cinder blueprint for re-imaging a volume, see more detail information in References.
The following tests are added.
Nova unit tests for negative scenarios
Nova functional tests for “happy path” testing
Tempest integration tests to make sure the nova/cinder integration works properly
We will replace the note in the API reference with a note about the required minimum microversion for rebuilding a volume-backed server with a new image.
The following document will be updated:
Stein PTG etherpad: https://etherpad.openstack.org/p/nova-ptg-stein
This is the discussion about rebuild the volume-backed server:
This is the discussion about what we should do about the root volume during a rebuild:
The cinder blueprint for re-imaging a volume: