[openstack-dev] [Neutron] Flavor framework proposal

Stephen Balukoff sbalukoff at bluebox.net
Thu Jul 17 03:36:26 UTC 2014


Hi Salvatore!

Thank you for reading through my book-length e-mail and responding to all
my points!

Unfortunately, I have more responses for you, inline:

On Wed, Jul 16, 2014 at 4:22 PM, Salvatore Orlando <sorlando at nicira.com>
wrote:

> Hi Stephen,
>
> Thanks for your exhaustive comments!
>

I'm always happy to exhaust others with my comments. ;)


> I think your points are true and valid for most cloud operators; besides
> the first all the point you provided indeed pertain operators and vendors.
> However you can't prove, I think, the opposite - that is to say that no
> cloud operator will find multi-service flavors useful. At the end of the
> day Openstack is always about choice - in this case the choice of having
> flavours spanning services or flavours limited to a single service.
> This discussion however will just end up slowly drifting into the realm of
> the theoretical and hypotethical and therefore won't bring anything good to
> our cause. Who know, in a few post we might just end up calling godwin's
> law!
>

That's certainly true.  But would you be willing to agree that both the
model and logic behind single-service_type flavors is likely to be simpler
to implement, troubleshoot, and maintain than multi-service_type flavors?

If you agree, then I would say: Let's go with single-service_type flavors
for now so that we can actually get an implementation done by Juno (and
thus free up development that is currently being blocked by lack of
flavors), and leave the more complicated multi-service_type flavors for
some later date when there's a more obvious need for them.

For what it's worth, I'm not against multi-service_type flavors if someone
can come up with a good usage scenario that is best solved using the same.
But I think it's more complication than we want or need right now, and
shooting for it now is likely to ensure we wouldn't get flavors in time for
Juno.



>  There are other considerations which could be made, but since they're
>>> dependent on features which do not yet exist (NFV, service insertion,
>>> chaining and steering) I think there is no point in arguing over it.
>>>
>>
>> Agreed. Though, I don't think single-service flavors paint us into a
>> corner here at all. Again, things get complicated enough when it comes to
>> service insertion, chaining, steering, etc. that what we'll really need at
>> that point is actual orchestration. Flavors alone will not solve these
>> problems, and orchestration can work with many single-service flavors to
>> provide the illusion of multi-service flavors.
>>
>
> Don't take it the wrong way - but this is what I mean by "theoretical and
> hypothetical". I agree with you. I think that's totally possible. But there
> are so many pieces which are yet missing from the puzzle that this
> discussion is probably worthless. Anyway, I started it, and I'm the one to
> be punished for it!
>

Hah! Indeed. Ok, I'll stop speculating down that path for now, eh. ;)


>  In conclusion I think the idea makes sense, and is a minor twist in the
>>> current design which should not either make the feature too complex neither
>>> prevent any other use case for which the flavours are being conceived. For
>>> the very same reason however, it is worth noting that this is surely not an
>>> aspect which will cause me or somebody else to put a veto on this work item.
>>>
>>
>> I don't think this is a minor twist in the current design, actually:
>> * We'll have to deal with cases like the above where no valid service
>> profiles can be found for a given kind of flavor (which we can avoid
>> entirely if a flavor can have service profiles valid for only one kind of
>> service).
>>
>
> Point taken, but does not require a major change to the design since a
> service flavour like this should probably be caught by a validation
> routine. Still you'd need more pervasive validation in different points of
> the API.
>

... which sounds like significantly more complication to me. But at this
point, we're arguing over what a "minor twist" is, which is not likely to
lead to anywhere useful...


>  * When and if tags/capabilities/extensions get introduced, we would need
>> to provide an additional capabilities list on the service profiles in order
>> to be able to select which service profiles provide the capabilities
>> requested.
>>
>
> Might be... but I don't see how that would be worse with multiple service
> types, especially if profiles are grouped by type.
>

Presumably, with single-service_type flavors, all service profiles
associated with the flavor should be capable of providing all the features
advertised as being provided by the flavor (first in the 'description' and
possibly later programmatically via an extensions list). This means we
don't have to check to see whether a service profile associated with the
flavor can provide for all the extensions advertised in the flavor
description because by creating the association, the operator is implying
it can.


>  * The above point makes things much more complicated when it comes to
>> scheduling algorithms for choosing which service profile to use when
>> multiple can meet the need for a given service. What does 'weight' mean if
>> all but two low-weight service profiles get eliminated as not suitable?
>>
>
> I'm really not following you. Sorry about my ignorance. Isn't this
> something you would need to address even in the case you have a flavour
> with multiple service profiles all for the same service type, as the spec
> currently mandates?
>

No, I'm probably simply not making my point clearly. What I mean is
probably best illustrated by an example:

Suppose I'm using service profile weights in my scheduling algorithm to
determine which service profile will get used when a user requests an
instance of an advanced service using the flavor framework.

For a single-service_type flavor, let's say I have the following service
profiles:

"Profile A"  => weight 30
"Profile B"  => weight 50
"Profile C"  => weight 20

I don't have to mention the type of service in the above profiles because
it's implied by the service_type attribute of the flavor. Given the above
weights, it's clear to see that if a user requests a service instance,
there's a 50% chance it'll be a "Profile B" instance. Then, later, when the
Operator looks at the relative split of services as deployed by users, she
should expect to see a "30 / 50 / 20" percentage split. A "5 / 90 / 5"
split would be a solid indication that "something is wrong" that the
operator needs to look into. (Maybe "Profile A" deployments fail silently
75% of the time, or maybe users have figured out that they hate "profile C"
instance types and keep deploying / destroying instances until they land on
a "profile B" instance.)

Now, doing the same exercise for a multi-service_type flavor, let's say we
have the following service profiles:

"Profile X" => service_types = [LBaaS, VPNaaS], weight = 30
"Profile Y" => service_types = [LBaaS, VPNaaS], weight = 50
"Profile Z" => service_types = [FWaaS, VPNaaS], weight = 20

Given the above:
* If a user tries to deploy an LBaaS instance, it's less obvious that the
instance will have a 62.5% chance of ending up on Profile Y, and a 37.5%
chance of ending up on Profile X.
* It's also less obvious that even though "Profile Z" has a low weight, if
the user wants a FWaaS instance, there's a 100% chance it'll end up on
Profile Z.
* From the Operator's perspective, it's no longer obvious that a "5 / 90 /
5" split of services on this flavor indicates "something is wrong." In
fact, it's much more difficult to determine this.

Does this example make sense to you?


>
>
>
>>
>> Another aspect to consider is how the flavours will work when the
>>> advanced service type they refer to is not consumable through the neutron
>>> API, which would be the case with an independent load balancing API
>>> endpoint. But this is probably another story.
>>>
>>
>> As far as I'm aware, flavors will only ever apply to advanced services
>> consumable through the Neutron API. If this weren't the case, what's the
>> point of having a flavor describing the service at all? If you're talking
>> about Octavia here--  well, our plan is to have Octavia essentially be an
>> other load balancer vendor, interfaced through a driver in the Neutron
>> LBaaS extension. (This is also why so many developers interested in seeing
>> Octavia come to light are spending all their time right now improving
>> Neutron LBaaS-- we want it to be feature-rich enough in Juno for us to be
>> able to actually do interesting things with Octavia without having to
>> resort to building our own independent API endpoint if at all possible.) If
>> Octavia ever splits off and has its own API endpoint, it would need to
>> implement something like the Neutron flavor framework itself-- and in this
>> case, what we decide to do there should not affect the Neutron flavor
>> framework at all (or visa versa).
>>
>
> This is a bit off topic, and again I am at fault for driving you there.
> The thing is that I've spoken with several people, and they've all got
> different opinions regarding where Load Balancing should go in the medium
> term. These range from a completely standalone service to what you've just
> mentioned here.
> My opinion, which is probably irrelevant, is that I'm supportive of the
> load balancing effort evolving independently from the rest of neutron and
> staying loosely coupled with it. However, I believe that there should a
> single API endpoint for all the network services. In other words a single
> neutron API offering all service to consumer and different standalone
> services providing those service at control/data plane.
>

Eh-- let's cross that bridge when we get there, eh. I think it will become
more obvious as projects like Octavia gain maturity whether they should be
split off and become completely independent, be loosely coupled, or simply
remain a "vendor" of Neutron LBaaS. :)  How "flavors" get handled in these
scenarios is part of that discussion-- but that discussion probably isn't
relevant right now.


>
>> If you want to provide the illusion of two different top-level services /
>> API endpoints having the same "flavor," then I would say, "that's what
>> orchestration is for."
>>
>
> Totally agree on this point.
>
>>
>> On Tue, Jul 15, 2014 at 2:07 PM, Eugene Nikanorov <
>> enikanorov at mirantis.com> wrote:
>>
>> Hi Stephen,
>>>
>>> So, as was discussed, existing proposal has some aspects which better to
>>> be postponed, like extension list on the flavor (instead of tags).
>>>
>>
>>  Agreed-- I think we need to more fully flesh out how extension list /
>> tags should work here before we implement it. But this doesn't prevent us
>> from rolling forward with a "version 1" of flavors so that we can start to
>> use some of the benefits of having flavors (like the ability to use
>> multiple service profiles with a single driver/provider, or multiple
>> service profiles for a single kind of service).
>>
>>
>>> Particularly that idea has several drawbacks:
>>>  - it makes public API inflexible
>>>  - turning features on/off is not what flavors should be doing, it's a
>>> task for policy framework and not flavors
>>>  - flavor-based rest call dispatching is quite complex solution giving
>>> no benefits for service plugins
>>>
>>
>> I'm confused as to what you mean by "that idea" here. Are you taking
>> about the "extension list"? If this is the case, I agree that that aspect
>> needs to be refined and should probably be postponed if possible.
>>
>
> I was under the impression that the extension list was a set of neutron
> extensions supported by all the service providers in the flavour. This
> could probably be enforced as an API level constraint.
>

I don't think there's much agreement on what the extension list should
actually be at the present time (honestly, it seem close enough to the
'tags' proposal in Eugene's original spec that I keep mistaking one for the
other).  This is one of the reasons I'm in favor of deferring that
discussion until version 2 of flavors, and simply working with the
free-form 'description' field for now which is informational but shouldn't
be considered programmatically consumable.


>
>>
>>> While this is not explicitly written in proposal - that's what implied
>>> there.
>>> I think that one is a major blocker of the proposal right now, it
>>> deserves future discussion and not essential to the problem flavors are
>>> supposed to solve.
>>>
>>
>> Yes, I think there are many benefits we can get out of the flavor
>> framework without having to have an extensions list / tags at this
>> revision. But I'm curious: Did we ever define what we were actually trying
>> to solve with flavors?  Maybe that's the reason the discussion on this has
>> been all of the place: People are probably making assumptions about the
>> problem we're trying to solve and we need to get on the same page about
>> this.
>>
>
> That is what I've been saying for over a year. Whatever you call it
> "service type", "service provider" or "flavour" it appears that's
> impossible to find two persons who think about it in the same way.
>

Haha! Great! Well, when it comes to proposals like this, I prefer to think
of them in terms of "what problem are we trying to solve" and then limiting
scope so we can actually produce something before the cold death of the
universe.

So! With that in mind, I agree with what Eugene said on this:

 The original problem has several aspects:
1) providing users with some information about what service implementation
they get (capabilities)
2) providing users with ability to specify (choose, actually) some
implementation details that don't relate to a logical configuration
(capacity, insertion mode, HA mode, resiliency, security standards, etc)
3) providing operators a way to setup different modes of one driver
4) providing operators to seamlessly change drivers backing up existing
logical configurations (now it's not so easy to do because logical config
is tightly coupled with provider/driver)

The proposal we're discussing right is mostly covering points (2), (3) and
(4) which is already a good thing.
So for now I'd propose to put 'information about service implementation' in
the description to cover (1)


If there are problems people are trying to solve with flavors that are not
among the above 4 points, I would suggest that these either become a part
of a later revision of flavors, or simply get discussed as a new entity
entirely (depending on what people are after).

Anyone have objections to this?

(Sorry for the abruptness on my part on this: We really need to get flavors
into Juno because we have features of LBaaS which are probably going to
depend on it. I realize I've been dominating this discussion because of the
time pressure here, but I really am interested in learning whether flavors
as proposed in the spec right now provide useful functionality to the other
Neutron Advanced Services, other than LBaaS.)



> Other than that, I personally don't have much disagreements on the
>>> proposal.
>>>
>>> The question about service type on the flavor is minor IMO. We can allow
>>> it to be NULL, which would mean multiservice flavor.
>>> However, multiservice flavors may put some minor requirements to driver
>>> API (that's mainly because of how flavor plugin interacts with service
>>> plugins)
>>>
>>
>> Yes, I think single-service flavors is almost certainly going to be a
>> simpler thing to implement, too-- and if we want to get flavors in for Juno
>> (which I know *we* really want to do), then I think reducing the complexity
>> here is probably a good idea, at least for the first revision.
>>
>
> Yes. At the end of the day I have no problem with going with single
> service flavours. Mostly because it will be already a win if we manage to
> bind this framework to a single advanced service in the juno release cycle.
>

Yay! That's the goal: Get this into Juno!  And it seems to me like a good
first step toward multi-service_type flavors, if we as a group decide
that's where we want to go.

Thanks,
Stephen

-- 
Stephen Balukoff
Blue Box Group, LLC
(800)613-4305 x807
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140716/8e85e198/attachment.html>


More information about the OpenStack-dev mailing list