Open Stack

On 3/29/2018 9:28 AM, Sean McGinnis wrote:
> I do not think changing the revert to snapshot implementation is appropriate
> here. There may be some cases where this can get the desired result, but there
> is no guarantee that there is a snapshot on the volume's base image state to
> revert to. It also would not make sense to overload this functionality to
> "revert to snapshot if you can, otherwise do all this other stuff instead."
> 

Agree.

> This would need to be a new API (microversioned) to add a reimage call. I
> wouldn't expect implementation to be too difficult as we already have that
> functionality for new volumes. We would just need to figure out the most
> appropriate way to take an already in-use volume, detach it, rewrite the image,
> then reattach it.

Agree.

> 
> Ideally, from my perspective, Nova would take care of the detach/attach portion
> and Cinder would only need to take care of imaging the volume.

Agree. :) And yeah, I pointed this out in the nova spec for 
volume-backed rebuild also. I think nova can basically handle this like 
it does for shelve today, and we'd do something like this:

1. disconnect the volume from the host
2. create a new empty volume attachment for the volume and instance - 
this is needed so the volume stays 'reserved' while we re-image it
3. delete the old volume attachment
4. call the new cinder re-image API
5. once the volume is available (TODO: how would we know?)
6. re-attach the volume by updating the attachment with the host 
connector, connect on the host, and complete the attachment (marks the 
volume as in-use again)

-- 

Thanks,

Matt