[Engine-devel] Fwd: Custom properties per device + vNIC profile = not working (< 3.3)

Mike Kolesnik mkolesni at redhat.com
Wed Nov 20 07:07:02 UTC 2013


----- Original Message -----
> On 11/11/2013 11:48 AM, Mike Kolesnik wrote:
> >
> > ----- Original Message -----
> >> Hi Mike,
> >>
> >> ----- Original Message -----
> >>> From: "Mike Kolesnik" <mkolesni at redhat.com>
> >>> To: "engine-devel" <engine-devel at ovirt.org>
> >>> Cc: "Barak Azulay" <bazulay at redhat.com>, "Martin Perina"
> >>> <mperina at redhat.com>, "Livnat Peer" <lpeer at redhat.com>,
> >>> "Itamar Heim" <iheim at redhat.com>
> >>> Sent: Monday, November 11, 2013 8:49:33 AM
> >>> Subject: Custom properties per device + vNIC profile = not working (<
> >>> 3.3)
> >>>
> >>> Hi,
> >>>
> >>> I came across a situation where I wanted to define custom properties on a
> >>> vNIC profile sitting under a network in a 3.2 data center.
> >>>  From what I saw the configuration value for custom properties
> >>> (CustomDeviceProperties) is split into 4, one per each version (3.0, 3.1,
> >>> 3.2, 3.3).
> >>> Since vNIC profile is located under the DC tree, it takes the DC version
> >>> -
> >>> 3.2 in this specific case.
> >>
> >> Custom Device Properties were designed to be specified for each cluster
> >> version
> >> independently, it doesn't care about DC version. AFAIK cluster version
> >> defines
> >> what features are available ...
> >>
> >>>
> >>> I tried to set the config value for 3.2 but got:
> >>> $ engine-config -s
> >>> CustomDeviceProperties="{type=interface;prop={myProp=[a-zA-Z0-9-]+}}"
> >>> --cver=3.2
> >>> Cannot set value {type=interface;prop={myProp=[a-zA-Z0-9-]+}} to key
> >>> CustomDeviceProperties. Device custom properties are not supported in
> >>> version 3.2
> >>>
> >>> This is already not very good, since in a 3.2 DC there can be 3.3
> >>> clusters
> >>> with 3.3 hosts that do support custom device properties.
> >>
> >> Specify your properties for 3.3 version, since they will be used in 3.3
> >> clusters ...
> >>
> >
> > But the effective version is the DC version as I explained.
> >
> > In a DC 3.0-3.3 I can have clusters which the minimal version is the DC
> > version, and the maximal version is 3.3.
> > For example I can have the following:
> > DC - version 3.0
> >   + Cluster 1 - version 3.0
> >   + Cluster 2 - version 3.1
> >   + Cluster 3 - version 3.2
> >   + Cluster 4 - version 3.3
> >
> > In this constellation, I could use custom device properties only on Cluster
> > 4, but it's not possible to define them since the vNIC profile is using
> > the DC version 3.0.
> > So effectively this feature is not usable to me unless I use a 3.3 DC.
> >
> >>>
> >>> I also tried to alter the config value in the DB directly, but the custom
> >>> properties code ignored it since custom properties are not supported in
> >>> 3.2.
> >>> So, de facto, I have no reasonable way as a user to define custom device
> >>> properties to use for my vNIC profiles in DC < 3.3.
> >>
> >> There are two configuration properties  for Custom Device Properties:
> >>
> >> 1) SupportCustomDeviceProperties
> >>       - defines in what version properties are supported
> >>       - cannot be altered by users of course
> >>
> >> 2) CustomDeviceProperties
> >>       - holds properties specification for each version
> >>       - can be defined using engine-config
> >>
> >>>
> >>> I opened the bug https://bugzilla.redhat.com/show_bug.cgi?id=1028757 for
> >>> this, however I also want to discuss the situation:
> >>
> >> I looked at the bug and the problem is, that management network profile
> >> is bound to DC and not the Cluster. And that's something we never thought
> >> of
> >> ...
> >>
> >>>
> >>> 1. As a user, I can't set custom properties for level < 3.3 which is not
> >>> good.
> >>
> >> Well, it's 3.3 feature, so it looks OK for me
> >>
> >>> Removing the blocking, and loading custom properties for all versions
> >>> would
> >>> fix the bug and allow using custom device properties for older versions,
> >>> the
> >>> reasonable place to block this would be running a VM (or plugging a
> >>> device).
> >>> Basically this is the lesser issue..
> >>>
> >>> 2. I just don't see the added value of splitting the definition of the
> >>> properties per level..
> >>
> >> The idea behind the version splitting was:
> >>
> >> 1) We have a device with a feature that doesn't work correctly with
> >> version
> >> 3.3,
> >>     but it's fixed in 3.4
> >> 2) By specifying custom property per version we cane disable this feature
> >> for
> >> 3.3
> >>     and enable for 3.4
> >
> > Custom properties is not for specifying which features are enabled, there
> > is a whole other mechanism for that..
> >
> > Custom properties is for hooks (and other possible extensions), which by
> > definition are not something that is guaranteed to exist so I see no point
> > to force the user to update multiple configurations and cause confusion
> > for him..
> 
> as martin explained, we have predefined custom properties, which are
> based on the vdsm version, and hence are actually features we know to
> expose or not to expose.
> user-defined custom properties - are up to the admin, but we let these
> be at cluster level as well to allow more granularity.

There are no predefined properties here, only user defined properties.

> 
> >
> >>
> >>> The custom properties are extensions which might or might not be
> >>> available
> >>> to
> >>> a certain VM, I don't see how having different sets of custom properties
> >>> per
> >>> version (what version, DC version, cluster version?) would make any
> >>> difference - either the VM can utilize the extension given some state of
> >>> the
> >>> system, or it can't, but the determining factor is not the version but
> >>> rather the availability of the extension.
> >>> For example, I can have a hook for vNIC altering some property installed
> >>> on
> >>> host A and not host B, if the VM runs on host A it will get this
> >>> capability
> >>> and on host B it won't, regardless the DC version the VM is in.
> >>>
> >>> This is not to say we shouldn't block custom properties on the
> >>> engine-VDSM
> >>> API level since it's only available since 3.3, but this is handled by
> >>> another config value (SupportCustomDeviceProperties) which is not
> >>> alterable
> >>> by the user.
> >>> So basically, I think splitting the value per version is over
> >>> complication
> >>> and see no added value to the users, just more maintenance should they
> >>> choose to use this feature.
> >>>
> >>> Your thoughts please.
> >>
> >> AFAIK only network and storage team wanted to use device custom properties
> >> in 3.3 version, but I'm not sure what's current usage status.
> >>
> >> But IMHO it's too late for 3.3 to change specification ...
> >
> > Since I can have cluster 3.3 in a DC < 3.3, and this is the upgrade path
> > for existing users,
> > I'd argue that the bug is severe enough and should be fixed asap even for
> > 3.3 versions.
> 
> please note that if you expose this at cluster level and not DC level,
> you need to make sure to verify it when moving a VM between clusters in
> same DC.
> also, if this is somehow related to logical networks, not vnic specific,
> than logical networks are DC level entities.
> 
> 

OK but my point was that a custom properties is not meant to be split by versions since
by definition of it, a hook might or might not exist on a given host (regardless of the host version).
It only imposes more strain on the user to define possible custom properties by version..

I see no value to users in this approach, only more work and unclearness..

Mind you, hook is not a "feature" that is explicitly supported on a given version, but an extension
mechanism which can have 3rd party extensions that might or might not exist on a given host, but this
won't stop an action from occurring (i.e. VM would still start if a hook is missing but some custom
property was sent).

Also the original bug still exists because even though the vNIC is sitting at VM which is in cluster
(thus in effect having access to the cluster version), the profile sits under network (which, as you
mention, is DC level entity).
So for the user using a DC < 3.3 there is no option to use this feature even though he can have 3.3
clusters in his DC.



More information about the Devel mailing list