4:15 p.m.
Hello all,
The shared raw disk feature description can be found under the following
links:
http://www.ovirt.org/wiki/Features/DetailedSharedRawDisk
http://www.ovirt.org/wiki/Features/SharedRawDisk
Please feel free, to share your comments.
Regards,
Maor
3:59 p.m.
On 02/02/2012 05:15 PM, Maor wrote:
1. Affected oVirt projects
i'm pretty sure the history data warehouse will need to adapt to this.
2. "The shared raw disk feature should provide the ability to attach
disk to many VMs with safe concurrent access,"
this could be read as if ovirt or vdsm somehow provides a mechanism for
safe concurrent access.
maybe something like "to multiple VMs that can handle concurrent access
to a shared disk without risk of corruption".
and having just written this - sounds like setting this flag at UI level
should include a prompt to the user to make sure they understand that
flagging the disk as shared *will* lead to corruption if it is attached
to virtual machines which do not support and expect it to be shared with
other virtual or physical machines[1]
3. "The synchronization/clustering of shared raw disk between VMs will
be managed in the file system. "
either i don't understand what this mean, or it could be read with a
misleading meaning.
4. VM Pools
VM Pools are always based (at least today) on templates, and templates
have no shared disks.
I'd just block attaching a shared disk to a VM which is part of a pool
(unless there is a very interesting use case meriting this)
5. "Quota has to be taken in consideration, for every new feature that
will involve consumption of resources managed by it."
I thought quota is not relevant in this feature.
6. future work - Permissions should be added for disk entity
so who can add a shared disk?
same as for floating disks, i find it hard to imagine a flow in which if
someone flagged a disk as shared, suddenly everyone can have access to it.
same as my statement of floating disks - I'll spend some more time to
reflect on this specific part.
[1] an external LUN based disk could be shared with a physical server as
well.
2:14 p.m.
On 02/03/2012 04:59 PM, Itamar Heim wrote:
I agree, I will change it.
Maybe the following rephrase will be more accurate: "The
synchronization/clustering of shared raw disk between VMs should be
based on external independent application which will be synchronized
with the guest application."
If there is no reason
to attach shared disk to a VM from pool, maybe its
also not that relevant to attach shared disk to stateless VM.
Miki?
Why not?
Quota should be taken in consideration when adding new shared raw disk,
or moving a disk to other domains. We should also notice that shared raw
disk should not consume more quota when sharing the disk with different VMs.
6. future work - Permissions should be added for disk entity
so who can add a shared disk?
Data Center Administrator or System Administrator
will be initialized
with permissions for creating shared raw disk, or changing shared disk
to be unshared.
Regarding attach/detach disks to/from VM, I was thinking that for phase
one we will count on the user VM permissions. If user will have
permissions to create new disks on the VM, he will also have permissions
to attach new shared raw disk to it.
3:40 p.m.
----- Original Message -----
I think there is such a use-case in clustered environments (DB cluster, for example), in
which you have several disks that are not shared (OS, applications, etc.), and several
disks that are shared (DB disks).
In this case, in order to create this clustered environment, it will be nice if you create
a template with the regular disks, create a pool from it, and attach all the VMs in the
pool the shared DB disks.
(It would be even nicer if the shared disk will be a part of the template, and when
creating VMs from it they will have have a "link" to this shared disk - but I
agree that it may be complex so maybe we should leave it aside for now).
Thoughts?
4:13 p.m.
On 02/05/2012 04:40 PM, Oved Ourfalli wrote:
Unless you have a correct 'uniqifying' process ('sysprep' for Windows),
this is really not going to work.
For example, even Windows sysprep fails at times. We've known for quite
some time that it does not make the DTC UUID unique, so when
cloning+sysprep'ing the system, you are still left with severe
connectivity issues at the DTC RPC level.
I therefore assume other applications may suffer from it as well (how do
you make a RHEVM instance unique? which fields in the DB, config files,
etc. do you have to re-create?).
Y.
5:59 p.m.
On 02/05/2012 02:14 PM, Maor wrote:
...
"The synchronization/clustering of shared
raw disk between VMs is the
responsibility of the guests. Unaware guests will lead to corruption of
the shared disk."
I think pools and stateless are different. I can envision a use case
where stateless guests would use a shared disk (say, in read only for
same data).
this means they can attach shared disks from other VMs they have no
permission on...
as i said earlier - need to think about this one some more.
11:35 a.m.
----- Original Message -----
Not sure what you mean here. I do think that if you have a pool of
Servers (rather than desktops),
It is a valid use case to be able have a shared raw disk attached to it.
I agree that this can be handled later on - but I would like the design at list to handle
it.
6:03 p.m.
On 02/02/12 17:15, Maor wrote:
Hi Maor,
- "when taking a VM snapshot, a snapshot of the shared disk will not be
taken."
I think it is worth mentioning that the shared disk will be part of the
VM snapshot configuration. The disk will appear as unplugged.
- Move VM is deprecated in 3.1.
- It seems from the wiki that shared disk is not supported for template
but is supported for VM pool.
I am not sure how can we do that? iirc we create pool from template.
What is the complexity of supporting shared disk in Templates? off the
top of my head it seems like it is more complicated to block shared
disks in templates than to support it. What do you think?
Livnat
11:06 p.m.
I started writing the changes in the email but got tired of it and just made a bunch of
changes in the wiki (impossible to track in email, such things should be done using
etherpad or something).
A few questions though:
plugged vs. enabled - I thought we converged on attached/detached and enabled/disabled and
not plugged/unplugged?
* Shared disks are attached with R/W permissions.
- What about enabling R/O ? (esp. for stateless/pools)
* Template disks should not be shared.
- Why not? (read only)
* When exporting a VM, only the disks which are not shared will be exported.
- Why is the above not treated the same as a snapshot? the configuration will reference
the shared disk as unplugged? (or will it and it's just not clear?)
I didn't touch stateless/pools but should be fixed to reflect comments on this
thread.
Is Remove shared disk and Delete shared disk the same thing? if so, why the dual
terminology?
I don't quite follow the logic determining which section is under functionality and
which under user experience. For example, why do you have a 'Delete shared disk'
section in the Ux section but not a 'Move shared disk' section (there is no shared
disk specific logic visible in the delete action UI).
* Disk name should be generated automatically based on the vm name and disk number in the
VM. Description will be empty.
* New disk should enforce the user to enter a name for the disk.
- Huh? the above 2 items seem like an oxymoron, but I may be missing something...
* Attach/Detach of a shared disk can be performed only when the VM is in status
'down'.
- Why? under the functionality section you clearly stated that attaching a disk will
result in it being attached but disabled...
----- Original Message -----
8:17 a.m.
On 13/02/12 00:06, Ayal Baron wrote:
Do we have ack from libvirt that attaching disk in r/o mode actually
works? (I know we opened a bug for testing this but i can't find the bug
- Haim?)
It is not the same case as snapshot. We don't have in the export domain
the shared image.
The first line is referring to upgrade flow (it is under the upgrade
section), the second line is the behavior going forward.
I agree this is confusing, Maor I suggest you remove the general
behavior from the upgrade section.
I agree with Ayal on this.
Attach/Detach a disk should be enabled regardless if the VM is running
or not, this applies to all disks not only to shared disk.
Maor, few more questions:
* "Regular disk can become a shared raw disk, by editing the existing
disk and marking the 'share disk' property type."
There are limitation to this, for example disk with snapshots can not be
shared (we support only shared raw disks etc.)
* "When removing a VM with shared disks attached to it, the shared disks
will not be deleted. "
If the shared disk is not attached to any other VM, why don't we delete it?
I think we should behave with it as any other disk.
Maybe going forward when deleting a VM we should ask if to remove the
disks as well. This logic can apply to shared disk in the same way, I
don't think we should have a special logic around this for shared disks.
* The VDSM owner is missing from the doc, and vdsm is missing from the
affected projects.
Livnat
8:59 a.m.
----- Original Message -----
Do we plan to support deleting disks in the disks tab? (as part of supporting
floating disks)
If not then we should delete the disk (as we won't be able to delete it after the last
VM using it is deleted....).
If we do, then I think we should indeed ask the user whether he would like to delete those
disks or not (in case it is the last VM).
Deleting it implicitly might be wrong, as these disks might contain data that the user
would like to attach to a new VM.
6:44 p.m.
On 02/12/2012 07:03 PM, Livnat Peer wrote:
Agreed, I changed it
to the following:
when taking a vm snapshot, a snapshot of the shared disk should not be
taken, although it will be part of the VM snapshot configuration and the
disk will appear as unplugged.
- Move VM is deprecated in 3.1.
Right, I removed this anecdote from the wiki.
What I was
thinking about, is that the administrator can take a VM from
the pool and attach it a shared disk, after the VM was created (for
testing).
The motivation for adding shared disk was that each entity that can be
added with a disk can also be added with shared disk.
Today, Administrator can add a disk to a VM from pool, which might be
wrong behaviour, so maybe its better not to support it...
Implementation wize it
might be less complex, the problem is the use
cases it raises, some of them which I'm thinking about are:
* If the disk will be deleted from the DC, should we remove it from the
template? or leave an indication in the template that there was a shared
disk there, maybe should not allow to delete the disk in the first
place, until it is unattached from the template?
* What do we want to do when creating a template from VM with shared
disk - Should User choose whether to create a template with/without the
shared disk?
Blocking shared disk from template means creating the template without
the shared disk, the implementation for it is to check if the disk is
shared or not.
I think that if GUI will support attaching shared disk to multiple VMs,
there is no strong use case for allowing adding shared disk to a template.
8:17 a.m.
On 13/02/12 19:44, Maor wrote:
Since template configuration is 'read-only' you can not change a disk to
be plugged or unplugged.
I would say you can not delete a disk that is part of a template
regardless if it is shared or not.
If a user is creating a template from VM the configuration should be
identical to the VM.
I am not sure what the above comment means but remember that we have API
users as well as UI.
I think that if we don't have a strong case for not supporting shared
disk in templates the default should be to support it.
>
> Livnat
>
10:44 a.m.
On 02/14/2012 09:17 AM, Livnat Peer wrote:
So in that case template with shared disk, will
block the user from
removing the shared disk from the DC.
Won't it will make the flow for the user a bit complicated.
User who wants to remove the shared disk, will need to remove the VM's
which are based on the template and then remove the template it self.
11:17 p.m.
----- Original Message -----
There should be no such thing as 'delete disk with snapshots'
When deleting a disk the only thing that should be deleted is topmost layer.
The disk does not stop to exist in the history of the VM and when reverting to an old
snapshot the disk should be there.
Shared disks have no snapshots so there is no issue in this sense, however, if a template
has a reference to the disk then deleting the disk would effectively modify the template
configuration(?) and iirc engine does not allow changing template configuration post
creation (but we had a very long thread on this already).
I think we're approaching this the wrong way.
There are 2 possible problems we're trying to solve here and having the original
shared disk as part of the template is the wrong solution for both.
The first problem is - user wants to attach the shared disk to all VMs derived from the
template - in this case the shared disk is *not* a part of the template and what is needed
is an automatic way to configure newly created VMs that would allow to attach the shared
disk.
The second problem is - user wants the data on the shared disk to be part of the template
-
This is the general case for template seeing as a template is a *copy* of the original VM
with stripped identity.
in this case what is needed is a *copy* of the data and the fact that the disk is shared
is coincidental.
What the above means is that when you create a template from a VM with a shared disk the
user should choose whether the operation would also create a new disk and copy the content
of the shared disk to it (it is the user's responsibility to make sure that the data
does not change while this operation is taking place, but we can help a little there) or
exclude the shared disk from the template.
9:27 p.m.
On 02/15/2012 12:20 PM, Miki Kenneth wrote:
...
my view is the use case of shared disk defined at template level is
nice, but can be done at a later phase. the higher priority would be to
actually support a shared disk, even if user needs to attach it to the
VMs they instantiated from the template.
so we need to remember the template use case, and not do things making
it more complex later, but can also do some validations preventing it if
it makes things more complex.
4801
days inactive
4814
days old
20 comments
7 participants
participants (7)
-
Ayal Baron -
Itamar Heim -
Livnat Peer -
Maor -
Miki Kenneth -
Oved Ourfalli -
Yaniv Kaul