[Engine-devel] API design and plan

[Adam Litke]

On Thu, Dec 08, 2011 at 04:56:17AM -0500, Ayal Baron wrote:
> 
> 
> ----- Original Message -----
> > On Tue, Dec 06, 2011 at 08:46:57AM -0600, Adam Litke wrote:
> > > On Tue, Dec 06, 2011 at 02:58:59PM +0200, Dan Kenigsberg wrote:
> > > > On Mon, Dec 05, 2011 at 11:34:18AM -0600, Adam Litke wrote:
> > > > > Hi everyone.  On today's VDSM call we discussed the
> > > > > requirements, design, and
> > > > > plan for updating the API to include support for QMF and
> > > > > single-host REST API.
> > > > > All members present arrived at a general consensus on the best
> > > > > way to design the
> > > > > next-generation API.  I have tried to capture this discussion
> > > > > in the oVirt wiki:
> > > > > 
> > > > >      http://ovirt.org/wiki/Vdsm_API
> > > > > 
> > > > > Please take a look at this page and let's discuss any changes
> > > > > that may be needed
> > > > > in order to adopt it as a working plan that we can begin to
> > > > > execute.  Thanks!
> > > > >
> > > > 
> > > > Very nice, I've fixed two bullets about the future of the
> > > > xml-rpc.
> > > 
> > > Thanks... Updates look good to me.
> > > 
> > > > I think that we are missing something here: how do we model
> > > > Vdsm-to-Vdsm
> > > > communication, in a binding-blind way? I'm less worried about the
> > > > storage-based mailbox used for lvextend requests: my problem is
> > > > with
> > > > migration command.
> > > 
> > > Ok, interesting...  Besides migration, are there other features
> > > (current or
> > > planned) that would involve P2P communication?  I want to ensure we
> > > consider the
> > > full problem space.
> > 
> > Well, I can imagine we would like a host in distress to migrate VMs
> > to
> > whomever can take them, without central management driving this
> > process.
> > (CAVE split brain)
> > 
> > At the momemt I cannot think of something that cannot be implemented
> > by
> > QMF events. Ayal?
> > 
> > > 
> > > > Currently, the implementation of the "migrate" verb includes
> > > > contacting
> > > > the remote Vdsm over xml-rpc before issuing the libvirt
> > > > migrateToURI2
> > > > command ('migrationCreate' verb).
> > > > 
> > > > A Vdsm user who choose to use the REST binding, is likely to want
> > > > this to
> > > > be implemented this using a REST request to the destination. This
> > > > means
> > > > that the implementation of Vdsm depends on the chosen binding.
> > > > 
> > > > The issue can be mitigating by requiring the binding level to
> > > > provide a
> > > > "callback" for migrationCreate (and any other future Vdsm->world
> > > > requests).
> > > > This would complicate the beautiful png at
> > > > http://ovirt.org/wiki/Vdsm_API#Design ... Does anyone have
> > > > another
> > > > suggestion?
> > > 
> > > Actually, I think you are blending the external API with vdsm
> > > internals.  As a
> > > management server or ovirt-engine, I don't care about the protocol
> > > that vdsm
> > > uses to contact the migration recipient.  As far as I am concerned
> > > this is a
> > > special case internal function call.  For that purpose, I think
> > > xmlrpc is
> > > perfectly well-suited to the task and should be used
> > > unconditionally, regardless
> > > of the bindings used to initiate the migration.
> > > 
> > > So I would propose that we modify the design such that we keep an
> > > extremely thin
> > > xmlrpc server active whose sole purpose is to service internal P2P
> > > requests.
> > 
> > Interesting. We could avoid even that, if we could register a
> > callback
> > with libvirt, so that destination libvirtd called destination Vdsm to
> > verify that all storage and networking resources are ready, before
> > executing qemu. DanPB, can something like that be done? (I guess it
> > is
> > not realistic since we may need to pass vdsm-specific data from
> > source
> > to dest, and libvirt is not supposed to be a general purpose
> > transport.)
> > 
> > Dan.
> 
> I don't understand the issue.  The whole point of the REST API is to be an
> easily consumable *single* node management API.  Once you start coordinating
> among different nodes then you need clustering and management (either
> distributed or centralized), in both cases it is fine to require having a bus
> in which case you have your method of communications between hosts to replace
> current xml-rpc.

Implicit in this statement is an assertion that live migration between two vdsm
instances will not be supported without orchestration from an ovirt-engine
instance.  I don't agree with placing such a limitation on vdsm since p2p
migration is already well-supported by the underlying components (libvirt and
qemu).

> Requiring an additional xml-rpc server sounds wrong to me.

The other option is to support a migrateCreate binding in REST and QMF.

-- 
Adam Litke <agl at us.ibm.com>
IBM Linux Technology Center