[openstack-dev] [nova] Versioned notifications... who cares about the version?
Sean Dague
sean at dague.net
Mon Nov 30 12:32:16 UTC 2015
On 11/24/2015 10:09 AM, John Garbutt wrote:
> On 24 November 2015 at 15:00, Balázs Gibizer
> <balazs.gibizer at ericsson.com> wrote:
>>> From: Andrew Laski [mailto:andrew at lascii.com]
>>> Sent: November 24, 2015 15:35
>>> On 11/24/15 at 10:26am, Balázs Gibizer wrote:
>>>>> From: Ryan Rossiter [mailto:rlrossit at linux.vnet.ibm.com]
>>>>> Sent: November 23, 2015 22:33
>>>>> On 11/23/2015 2:23 PM, Andrew Laski wrote:
>>>>>> On 11/23/15 at 04:43pm, Balázs Gibizer wrote:
>>>>>>>> From: Andrew Laski [mailto:andrew at lascii.com]
>>>>>>>> Sent: November 23, 2015 17:03
>>>>>>>>
>>>>>>>> On 11/23/15 at 08:54am, Ryan Rossiter wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 11/23/2015 5:33 AM, John Garbutt wrote:
>>>>>>>>>> On 20 November 2015 at 09:37, Balázs Gibizer
>>>>>>>>>> <balazs.gibizer at ericsson.com> wrote:
>>>>>>>>>>> <snip>
>>>>>>>>>>> <snip>
>>>>>>>>>
>>>>>>>>>> There is a bit I am conflicted/worried about, and thats when we
>>>>>>>>>> start including verbatim, DB objects into the notifications. At
>>>>>>>>>> least you can now quickly detect if that blob is something
>>>>>>>>>> compatible with your current parsing code. My preference is
>>>>>>>>>> really to keep the Notifications as a totally separate object
>>>>>>>>>> tree, but I am sure there are many cases where that ends up
>>>>>>>>>> being seemingly stupid duplicate work. I am not expressing this
>>>>>>>>>> well in text form :(
>>>>>>>>> Are you saying we don't want to be willy-nilly tossing DB
>>>>>>>>> objects across the wire? Yeah that was part of the rug-pulling
>>>>>>>>> of just having the payload contain an object. We're
>>>>>>>>> automatically tossing everything with the object then, whether
>>>>>>>>> or not some of that was supposed to be a secret. We could add
>>>>>>>>> some sort of property to the field like
>>>>>>>>> dont_put_me_on_the_wire=True (or I guess a
>>>>>>>>> notification_ready() function that helps an object sanitize
>>>>>>>>> itself?) that the notifications will look at to know if it puts
>>>>>>>>> that on the wire-serialized dict, but that's adding a lot more
>>>>>>>>> complexity and work to a pile that's already growing rapidly.
>>>>>>>>
>>>>>>>> I don't want to be tossing db objects across the wire. But I
>>>>>>>> also am not convinced that we should be tossing the current
>>>>>>>> objects over the wire either.
>>>>>>>> You make the point that there may be things in the object that
>>>>>>>> shouldn't be exposed, and I think object version bumps is another
>>>>>>>> thing to watch out for.
>>>>>>>> So far the only object that has been bumped is Instance but in
>>>>>>>> doing so no notifications needed to change. I think if we just
>>>>>>>> put objects into notifications we're coupling the notification
>>>>>>>> versions to db or RPC changes unnecessarily. Some times they'll
>>>>>>>> move together but other times, like moving flavor into
>>>>>>>> instance_extra, there's no reason to bump notifications.
>>>>>>>
>>>>>>>
>>>>>>> Sanitizing existing versioned objects before putting them to the
>>>>>>> wire is not hard to do.
>>>>>>> You can see an example of doing it in
>>>>>>> https://review.openstack.org/#/c/245678/8/nova/objects/service.py,
>>>>>>> cm
>>>>>>> L382.
>>>>>>> We don't need extra effort to take care of minor version bumps
>>>>>>> because that does not break a well written consumer. We do have to
>>>>>>> take care of the major version bumps but that is a rare event and
>>>>>>> therefore can be handled one by one in a way John suggested, by
>>>>>>> keep sending the previous major version for a while too.
>>>>>>
>>>>>> That review is doing much of what I was suggesting. There is a
>>>>>> separate notification and payload object. The issue I have is that
>>>>>> within the ServiceStatusPayload the raw Service object and version
>>>>>> is being dumped, with the filter you point out. But I don't think
>>>>>> that consumers really care about tracking Service object versions
>>>>>> and dealing with compatibility there, it would be easier for them
>>>>>> to track the ServiceStatusPayload version which can remain
>>>>>> relatively stable even if Service is changing to adapt to db/RPC changes.
>>>>> Not only do they not really care about tracking the Service object
>>>>> versions, they probably also don't care about what's in that filter list.
>>>>>
>>>>> But I think you're getting on the right track as to where this needs
>>>>> to go. We can integrate the filtering into the versioning of the payload.
>>>>> But instead of a blacklist, we turn the filter into a white list. If
>>>>> the underlying object adds a new field that we don't want/care if
>>>>> people know about, the payload version doesn't have to change. But if
>>>>> we add something (or if we're changing the existing fields) that we
>>>>> want to expose, we then assert that we need to update the version of
>>>>> the payload, so the consumer can look at the payload and say "oh, in
>>>>> 1.x, now I get _______" and can add the appropriate checks/compat.
>>>>> Granted with this you can get into rebase nightmares ([1] still
>>>>> haunts me in my sleep), but I don't see us frantically changing the
>>>>> exposed fields all too often. This way gives us some form of
>>>>> pseudo-pinning of the subobject. Heck, in this method, we could even
>>>>> pass the whitelist on the wire right? That way we tell the consumer
>>> explicitly what's available to them (kinda like a fake schema).
>>>>
>>>> I think see your point, and it seems like a good way forward. Let's
>>>> turn the black list to a white list. Now I'm thinking about creating a
>>>> new Field type something like WhiteListedObjectField which get a type
>>>> name (as the ObjectField) but also get a white_list that describes which
>>> fields needs to be used from the original type.
>>>> Then this new field serializes only the white listed fields from the
>>>> original type and only forces a version bump on the parent object if
>>>> one of the white_listed field changed or a new field added to the
>>> white_list.
>>>> What it does not solve out of the box is the transitive dependency. If
>>>> today we Have an o.vo object having a filed to another o.vo object and
>>>> we want to put the first object into a notification payload but want to
>>>> white_list fields from the second o.vo then our white list needs to be
>>>> able to handle not just first level fields but subfields too. I guess
>>>> this is doable but I'm wondering if we can avoid inventing a syntax
>>> expressing something like 'field.subfield.subsubfield'
>>>> in the white list.
>>>
>>> Rather than a whitelist/blacklist why not just define the schema of the
>>> notification within the notification object and then have the object code
>>> handle pulling the appropriate fields, converting formats if necessary, from
>>> contained objects. Something like:
>>>
>>> class ServicePayloadObject(NovaObject):
>>> SCHEMA = {'host': ('service', 'host'),
>>> 'binary': ('service', 'binary'),
>>> 'compute_node_foo': ('compute_node', 'foo'),
>>> }
>>>
>>> fields = {
>>> 'service': fields.ObjectField('Service'),
>>> 'compute_node': fields.ObjectField('ComputeNode'),
>>> }
>>>
>>> def populate_schema(self):
>>> self.compute_node = self.service.compute_node
>>> notification = {}
>>> for key, (obj, field) in schema.iteritems():
>>> notification[key] = getattr(getattr(self, obj), field)
>>>
>>> Then object changes have no effect on the notifications unless there's a
>>> major version bump in which case a SCHEMA_VNEXT could be defined if
>>> necessary.
>>
>> Nice idea I will try it. Thanks! It is seems to avoid the sub object field white lists
>> problem as the needed notification field can always be pulled directly from an object field.
>
> +1
> This is my preference, specific notification objects that are
> independently versioned.
> It feels like time saving to re-use existing objects, but it breaks
> the interface really.
Ok, so that means we'll now have:
* REST representation (strongly versioned / documented)
* Notification representation (strongly versioned / documented )
* Nova objects representation (strongly versioned / documented only in code)
* Nova db objects (versioned by schema, documentated only in code)
If the notifications are not going to be raw Nova objects, I think we
need to really think about why they aren't the REST objects. Having a
whole other additional interface surface seems really really weird.
-Sean
--
Sean Dague
http://dague.net
More information about the OpenStack-dev
mailing list