[openstack-dev] [Nova] The unbearable lightness of specs

Matt Riedemann mriedem at linux.vnet.ibm.com
Wed Jun 24 13:25:49 UTC 2015



On 6/24/2015 7:42 AM, Daniel P. Berrange wrote:
> On Wed, Jun 24, 2015 at 11:28:59AM +0100, Nikola Đipanov wrote:
>> Hey Nova,
>>
>> I'll cut to the chase and keep this email short for brevity and clarity:
>>
>> Specs don't work! They do nothing to facilitate good design happening,
>> if anything they prevent it. The process layered on top with only a
>> minority (!) of cores being able to approve them, yet they are a prereq
>> of getting any work done, makes sure that the absolute minimum that
>> people can get away with will be proposed. This in turn goes and
>> guarantees that no good design collaboration will happen. To add insult
>> to injury, Gerrit and our spec template are a horrible tool for
>> discussing design. Also the spec format itself works for only a small
>> subset of design problems Nova development is faced with.
>
> I'd like to see some actual evidence to backup a sweeping statement
> as "Specs dont work. They do nothing to facilitate good design happening,
> if anything they prevent it."
>
> Comparing Nova today, with Nova before specs were introduced, I think
> that specs have had a massive positive impact on the amount of good
> design and critique that is happening.
>
> Before specs, the average blueprint had no more than 3 lines of text
> in its description. Occassionally a blueprint would link to a wiki
> page or google doc with some design information, but that was very
> much the exception.

Yeah, seems we've forgotten how before we had specs in gerrit everyone 
complained about the lack of detail and ability to add comments in the 
launchpad blueprint whiteboard. Storyboard was meant to be the savior 
but that took too long so specs in gerrit was going to be an alternative 
for awhile and it ended up working nicely.

Now that we're tired of the new thing we can start criticizing it and 
iterating on how to make it more agile.

>
> When I was reviewing features in Nova before specs came along, I spent
> alot of time just trying to figure out what on earth the code was
> actually attempting to address, because there was rarely any statement
> of the problem being addressed, or any explanation of the design that
> motivated the code.  This made life hard for reviewers trying to figure
> out if the code was acceptable to merge.  It is pretty bad for contributors
> trying to implement new features too, as they could spend weeks or months
> writing and submitting code, only to be rejected at the end because the
> (lack of any design discussions) meant they missed some aspect of the
> problem which in turn meant all their work was in vain. That was a collosal
> waste of everyone's time and resulted in some of the very horrible code
> impl decisions we're still living with today.
>
>> That's only a subset of problems. Some more, you ask? OK. No clear
>> guidelines as to what needs a spec, that defaults to "everything does".
>> And spec being the absolute worst way to judge the validity of some of
>> the things that do require them.
>>
>> Examples of the above are everywhere if you care to look for them but
>> some that I've hit _this week_ are [1] (spec for a quick and dirty fix?!
>> really?!) [2] (spec stuck waiting for a single person to comment
>> something that is an implementation detail, and to make matter worse the
>> spec is for a bug fix) [3] (see how ill suited the format is for a
>> discussion + complaints about grammar and spelling instead of actual
>> point being made).
>>
>> Nova's problem is not that it's big, it's that it's big _and_ tightly
>> coupled. This means no one can be trusted to navigate the mess
>> successfully, so we add the process to stop them. What we should be
>> doing is fixing the mess, and the very process is preventing that.
>>
>> Don't take my word for it - ask the Gantt subteam who have been trying
>> to figure out the scheduler interface for almost 4 cycles now. Folks
>> doing Cells might have a comment on this too.
>>
>> The good news is that we have a lot of stuff in place already to help us
>> reduce this massive coupling of everything. We have versioned objects,
>> we have versioned RPC. Our internal APIs are terrible, and provide no
>> isolation, but we have the means to iterate and figure it out.
>>
>> I don't expect this process issues will get solved quickly, If it were
>> up to me I'd drop the whole thing, but I understand that it's not how
>> it's done.
>>
>> I do hope this makes people think, discuss and move things into the
>> direction of facilitating quality software development instead of
>> outright preventing it. I'll follow up with some ideas on how to go
>> forward once a few people have commented back.
>
> I will agree that the specs process has a number of flaws - in particular
> I think we've treated it as too much of a rigid process resulting it in
> being very beurocractic. In particular I think are missing an ability to
> be pragmmatic in decisions about the level of design required and whether
> specs are required. The idea of allowing blueprints without specs in
> some cases was an attempt to address this, but I don't feel it has been
> very successful - there is still too much stuff being forced through
> the specs process unncessarily imho.
>
> I've repeatedly stated that the fact that we created an even smaller
> clique of people to approve specs (nova-drivers which is a tiny subset
> of the already faaaar too small nova-core) is madness, as it creates
> an even worse review burden on them, and thus worsens the bottleneck
> than we already have.
>
> In reviewing specs, we also often get into far too much detail about
> the actual implementation, which is really counterproductive. eg when
> get down to bikeshedding about the names of classes, types of object
> attributes, and so forth, it is really insane. That level of detail
> and bikesheeding belongs in the code review, not the design for the
> most part.
>
> Specs are also inflexible when it comes to dealing with features that
> cross multiple projects, because we silo'd spec reviews against the
> individual projects. This is sub-optimal, but ultimately not a show
> stopper.

There is the openstack-specs repo for cross-project specs:

http://specs.openstack.org/openstack/openstack-specs/

That'd be a good place for things like your os-vif library spec which 
requires input from both nova and neutron teams, although I think it's 
currently OK to have it in nova-specs since a lot of the forklift is 
going to come from there and we can add neutron reviewers as needed.

>
> I also think the way we couple spec approval & reviews to the dev
> cycles is counterproductive. We should be willing to accept and
> review specs at any point in any cycle, and once approved they should
> remain valid for a prolonged period of time - not require us to go
> through re-review every new dev cycle as again that's just creating
> extra burden. We should of course though reserve the right to unapprove
> specs if circumstances change, invlidating the design for the previous
> approval.
>
> In short specs are far from perfect, but to say specs don't solve/help
> anything todo with design in nova is really ignoring our history from
> before the time specs existed. We must continue to improve our process
> overall the biggest thing we lost IMHO is agility and pragmatism in
> our decision making - I think we can regain that without throwing away
> the specs idea entirely.
>
> Regards,
> Daniel
>

-- 

Thanks,

Matt Riedemann




More information about the OpenStack-dev mailing list