Open Stack

On Tue, Sep 18, 2018 at 03:07:45PM +0200, Attila Fazekas wrote:
> On Tue, Sep 18, 2018 at 2:09 PM, Peter Penchev <openstack-dev at storpool.com>
> wrote:
> 
> > On Tue, Sep 18, 2018 at 11:32:37AM +0200, Attila Fazekas wrote:
> > [format recovered; top-posting after an inline reply looks confusing]
> > > On Mon, Sep 17, 2018 at 11:43 PM, Jay Pipes <jaypipes at gmail.com> wrote:
> > >
> > > > On 09/17/2018 09:39 AM, Peter Penchev wrote:
> > > >
> > > >> Hi,
> > > >>
> > > >> So here's a possibly stupid question - or rather, a series of such :)
> > > >> Let's say a company has two (or five, or a hundred) datacenters in
> > > >> geographically different locations and wants to deploy OpenStack in
> > both.
> > > >> What would be a deployment scenario that would allow relatively easy
> > > >> migration (cold, not live) of instances from one datacenter to
> > another?
> > > >>
> > > >> My understanding is that for servers located far away from one another
> > > >> regions would be a better metaphor than availability zones, if only
> > > >> because it would be faster for the various storage, compute, etc.
> > > >> services to communicate with each other for the common case of doing
> > > >> actions within the same datacenter.  Is this understanding wrong - is
> > it
> > > >> considered all right for groups of servers located in far away places
> > to
> > > >> be treated as different availability zones in the same cluster?
> > > >>
> > > >> If the groups of servers are put in different regions, though, this
> > > >> brings me to the real question: how can an instance be migrated across
> > > >> regions?  Note that the instance will almost certainly have some
> > > >> shared-storage volume attached, and assume (not quite the common case,
> > > >> but still) that the underlying shared storage technology can be taught
> > > >> about another storage cluster in another location and can transfer
> > > >> volumes and snapshots to remote clusters.  From what I've found, there
> > > >> are three basic ways:
> > > >>
> > > >> - do it pretty much by hand: create snapshots of the volumes used in
> > > >>    the underlying storage system, transfer them to the other storage
> > > >>    cluster, then tell the Cinder volume driver to manage them, and
> > spawn
> > > >>    an instance with the newly-managed newly-transferred volumes
> > > >>
> > > >
> > > > Yes, this is a perfectly reasonable solution. In fact, when I was at
> > AT&T,
> > > > this was basically how we allowed tenants to spin up instances in
> > multiple
> > > > regions: snapshot the instance, it gets stored in the Swift storage
> > for the
> > > > region, tenant starts the instance in a different region, and Nova
> > pulls
> > > > the image from the Swift storage in the other region. It's slow the
> > first
> > > > time it's launched in the new region, of course, since the bits need
> > to be
> > > > pulled from the other region's Swift storage, but after that, local
> > image
> > > > caching speeds things up quite a bit.
> > > >
> > > > This isn't migration, though. Namely, the tenant doesn't keep their
> > > > instance ID, their instance's IP addresses, or anything like that.
> >
> > Right, sorry, I should have clarified that what we're interested in is
> > technically creating a new instance with the same disk contents, so
> > that's fine.  Thanks for confirming that there is not a simpler way that
> > I've missed, I guess :)
> >
> > > > I've heard some users care about that stuff, unfortunately, which is
> > why
> > > > we have shelve [offload]. There's absolutely no way to perform a
> > > > cross-region migration that keeps the instance ID and instance IP
> > addresses.
> > > >
> > > > - use Cinder to backup the volumes from one region, then restore them
> > to
> > > >>    the other; if this is combined with a storage-specific Cinder
> > backup
> > > >>    driver that knows that "backing up" is "creating a snapshot" and
> > > >>    "restoring to the other region" is "transferring that snapshot to
> > the
> > > >>    remote storage cluster", it seems to be the easiest way forward
> > (once
> > > >>    the Cinder backup driver has been written)
> > > >>
> > > >
> > > > Still won't have the same instance ID and IP address, which is what
> > > > certain users tend to complain about needing with move operations.
> > > >
> > > > - use Nova's "server image create" command, transfer the resulting
> > > >>    Glance image somehow (possibly by downloading it from the Glance
> > > >>    storage in one region and simulateneously uploading it to the
> > Glance
> > > >>    instance in the other), then spawn an instance off that image
> > > >>
> > > >
> > > > Still won't have the same instance ID and IP address :)
> > > >
> > > > Best,
> > > > -jay
> > > >
> > > > The "server image create" approach seems to be the simplest one,
> > > >> although it is a bit hard to imagine how it would work without
> > > >> transferring data unnecessarily (the online articles I've seen
> > > >> advocating it seem to imply that a Nova instance in a region cannot be
> > > >> spawned off a Glance image in another region, so there will need to be
> > > >> at least one set of "download the image and upload it to the other
> > > >> side", even if the volume-to-image and image-to-volume transfers are
> > > >> instantaneous, e.g. using glance-cinderclient).  However, when I tried
> > > >> it with a Nova instance backed by a StorPool volume (no ephemeral
> > image
> > > >> at all), the Glance image was zero bytes in length and only its
> > metadata
> > > >> contained some information about a volume snapshot created at that
> > > >> point, so this seems once again to go back to options 1 and 2 for the
> > > >> different ways to transfer a Cinder volume or snapshot to the other
> > > >> region.  Or have I missed something, is there a way to get the "server
> > > >> image create / image download / image create" route to handle volumes
> > > >> attached to the instance?
> > > >>
> > > >> So... have I missed something else, too, or are these the options for
> > > >> transferring a Nova instance between two distant locations?
> > > >>
> > > >> Thanks for reading this far, and thanks in advance for your help!
> > > >>
> > > >> Best regards,
> > > >> Peter
> > >
> > > Create a volume transfer VM/machine in each region.
> > > attache the volume -> dd -> compress  -> internet   ->decompress -> new
> > > volume, attache(/boot with) to the volume to the final machine.
> > > In case you have frequent transfers you may keep up the machines for the
> > > next one..
> >
> > Thanks for the advice, but this would involve transferring *a lot* more
> > data than if we leave it to the underlying storage :)  As I mentioned,
> > the underlying storage can be taught about remote clusters and can be told
> > to create a remote snapshot of a volume; this will be the base on which
> > we will write our Cinder backup driver.  So both my options 1 (do it "by
> > hand" with the underlying storage) and 2 (cinder volume backup/restore)
> > would be preferable.
> >
> 
> Cinder might get a feature for `rescue` a volume in case accidentally
> someone
> deleted the DB record or some other bad thing happened.
> This needs to be admin only op where you would need to specify where is the
> volume,
> If just a new volume `shows up` on the storage, but  without
> the knowledge of cinder, it could be rescued as well.

Hmm, is this not what the Cinder "manage" command does?

> Among same storage types probably cinder could have an admin only
> API for transfer.
> 
> I am not sure is volume backup/restore is really better across regions
> than the above steps properly piped however
> it is very infrastructure dependent,
> bandwidth and latency across regions matters.
[snip discussion]

Well, the reason my initial message said "assume the underlying storage
can do that" was that I did not want to go into marketing/advertisement
territory and say flat out that the StorPool storage system can do that :)

Best regards,
Peter

-- 
Peter Penchev  openstack-dev at storpool.com  https://storpool.com/