Few questions on Backup Restore API implmn.

[Michael Pasternak]

Hi Deepak,

On 05/30/2013 06:07 PM, Deepak C Shetty wrote:
> Hi All,
>     Per the previous mail from Ayal.. I am refering to the flow below...
> 
> The backup flow should be (all api calls are engine REST API):

I disagree with this, by following this pattern you move internal
logic (even more than this given implementation specific logic) to
the api users instead of abstracting it and creating solid api for
backups,

+++++++++ several api calls for single backup ++++++++++++++


cons:
====

1. user have to be aware of internal (implementation specific) flow of each provider in order to be able creating backup
   1.1 call add(a)->do(a)->del(a)->do(b)->etc.
2. no flexibility for backup provider to change the backup flow (as it would be considered an api break)
3. no atomic resolution for the /backup failures
   3.1. failure at do(b) of add(a)->do(a)->del(a)->do(b)->etc. will mean
        that user have to del(a) manually and if it's deeper in the flow
        every api user will have to implement own rollbacks to the steps that took
        place before the failure.
4. backward compatibility becomes an issue when exposing internals in such a low level.
5. new features expose becomes much complex task cause they may require performing extra
   step/s in a middle.

pros:
====

can't think of any ..., cause it doesn't give you flexibility, but creates unnatural
(to public APIs) complexity.


+++++++++ single api call for single backup ++++++++++++++

cons:
====

nothing comes to my mind, are there any?

pros:
====

1. user is not aware of any backup internals and can consume different /backup
   providers using very same api and supplying parameters relevant for the provider
   he'd like to use.
2. backup providers can change/extend internal flow as much as they want/need, only
   by modifying internal engine driver, user using abstraction over the backup api
   won't feel the difference.
3. backup providers can execute backup as single atomic operation and take care
   for rollback cleanups in the transaction.
4. backup providers can easily maintain backward compatibility over abstraction
5. adding new features that require flow extension can be easily hidden under
   the abstraction layer.

also this is a common practice to expose abstraction layer to public, while
vendors only have to implement the driver in order to be supported.


> 1. create vm snapshot (the api call exists today)
> 2. get VM Config (new API)
> 3. prepare backup disk (new api, should probably accept: hostid, disk;
> return: paths to device on host as well as map of changed blocks) -
> this should be called for every disk that needs to be backed up. Note
> that VM snapshot takes a snap of *all* disks of the VM
> 4. attach disk to backup vm (the api call exists today.  This assumes
> virt app) - also has to be called per disk to back up
> 5. At this point backup app would do the backup
> 

this can be easily done under the single api call:
=================================================

POST /api/vms/xxx/disks/yyy/backups

<backup>
  <host id=aaa> #3
  <backup_vm id=bbb> #1, #2
  <delete_snapshot>true|false<delete_snapshot/> #7
<backup/>

1. create vm snapshot => you already in context of vm, trigger snapshot on it
2. get VM Config => you already in context of vm, collect relevant meta
3. prepare backup disk => you in context of the disk already, do whatever is needed
4. attach disk to backup vm => you have backup_vm id in the request body
5. do the actual backup

does this makes sense?

> 
> 5) detach disk (symmetric to 4)
> 6. teardown backup disk (symmetric to 3)
> 7. delete snap - This can only be called if VM is down today and is
> not mandatory in any event.  Once we have live merge, it should be
> policy driven (user should be able to choose whether to keep or delete)
> 
> Questions...
> 
> 1) My current implmn of prepareBackupDisk in VDSM (pls see http://gerrit.ovirt.org/#/q/status:open+project:vdsm+branch:master+topic:backup-restore,n,z)
> returns drive and cbtinfo as dicts back to engine. The understnading here is that the drive returned will be a dict which has all the necessary info
> to represent a NBD drive to libvirt... hence this drive will be passed as-is in the "attach disk to VM" REST API. Is this correct ?
> 
> Note that i cannot return a disk UUID (like VDSM does for create volume case).. because in preparebackupdisk case the image is exported usign qemu-nbd
> as a block device over the network... and hence its not a regular vdsm-type disk, hence no UUID, Agree ?
> 
> 2) In "attach disk to VM" REST API user will pass the drive dict which he/she got as part of preparebackupdisk api.. as-is
> and VDSM will need to add support for accepting NBD disk as a valid disktype in VM.hotplugDisk() code of VDSM.
> 
> The ability to add a network disk is present in libvirtvm.py... as part of my GlusterSD work, sicne gluster is represented as a NBD to QEMU
> but it doesn't work at the API.py level.. its only from prepareImage onwards... hence the need to modify API.py to accept NBD disk type and connect
> to the libvirtvm.py appropriately
> 
> Is this acceptable.. otherwise there is no good way to pass the NBD drive's info back to VDSM as part of existing "attach disk to VM" API
> Also does the existing "attach disk to VM" API work if a pre-loaded drive dict/hashmap is provided by the user. This will have drive->type = network
> and not file/blcok as is currently supported
> 
> 3) After "attach disk to backup VM" REST API step.. the understnading is that some backup vendor specific API will be called to tell the
> backup appln....
>     -- Block device (attached to the VM) to be used as src for backup
>     -- CBT info of this blcok device (which was recd. as part of prepareBackupDisk API)
> Is this correct ?
> 
> thanx,
> deepak
> 


-- 

Michael Pasternak
RedHat, ENG-Virtualization R&D