Open Stack

2016-02-12 11:19 GMT-08:00 Andrew Laski <andrew at lascii.com>:
> On Fri, Feb 12, 2016, at 01:45 PM, John Garbutt wrote:
>> On 12 February 2016 at 18:17, Andrew Laski <andrew at lascii.com> wrote:
>> >
>> >
>> > On Fri, Feb 12, 2016, at 12:15 PM, Matt Riedemann wrote:
>> >> Forgive me for thinking out loud, but I'm trying to sort out how nova
>> >> would use a microversion in the nova API for the get-me-a-network
>> >> feature recently added to neutron [1] and planned to be leveraged in
>> >> nova (there isn't a spec yet for nova, I'm trying to sort this out for a
>> >> draft).
>> >>
>> >> Originally I was thinking that a network is required for nova boot, so
>> >> we'd simply check for a microversion and allow not specifying a network,
>> >> easy peasy.
>> >>
>> >> Turns out you can boot an instance in nova (with neutron as the network
>> >> backend) without a network. All you get is a measly debug log message in
>> >> the compute logs [2]. That's kind of useless though and seems silly.
>> >>
>> >> I haven't tested this out yet to confirm, but I suspect that if you
>> >> create a nova instance w/o a network, you can latter try to attach a
>> >> network using the os-attach-interfaces API as long as you either provide
>> >> a network ID *or* there is a public shared network or the tenant has a
>> >> network at that point (nova looks those up if a specific network ID
>> >> isn't provided).
>> >>
>> >> The high-level plan for get-me-a-network in nova was simply going to be
>> >> if the user tries to boot an instance and doesn't provide a network, and
>> >> there isn't a tenant network or public shared network to default to,
>> >> then nova would call neutron's new auto-allocated-topology API to get a
>> >> network. This, however, is a behavior change.
>> >>
>> >> So I guess the question now is how do we handle that behavior change in
>> >> the nova API?
>> >>
>> >> We could add an auto-create-net boolean to the boot server request which
>> >> would only be available in a microversion, then we could check that
>> >> boolean in the compute API when we're doing network validation.
>> >>
>> >
>> > I think a flag like this is the right approach. If it's currently valid
>> > to boot an instance without a network than there needs to be something
>> > to distinguish a request that wants a network created vs. a request that
>> > doesn't want a network.
>> >
>> > This is still hugely useful if all that's required from a user is to
>> > indicate that they would like a network, they still don't need to
>> > understand/provide details of the network.
>>
>> I was thinking a sort of opposite. Would this work?
>>
>> We add a new micro-version that does this:
>> * nothing specified: do the best we can to get a port created
>> (get-me-a-network, etc,), or fail if not possible
>> * --no-nics option (or similar) that says "please don't give me any nics"
>>
>> This means folks that don't want a network, reliably have a way to do
>> that. For everyone else, we do the same thing when using either
>> neutron or nova-network VLAN manager.
>
> I think this pushes our microversions meaning a bit further than
> intended. I don't think the API should flip behaviors simply based on a
> microversion.
>
> What about requiring nic information with the microversion? Make users
> indicate explicitly if they want a network or not and avoid changing a
> default behavior.

I agree with alaski's point here.
The default behavior changes can hurt API users easily even if bumping
a microversion because it is difficult to pay attentions for all
microversions' changes(we have already 22 microversions now) from
users' viewpoints.

Thanks
Ken Ohmichi