Open Stack

On 01/26/2016 03:46 AM, Steven Hardy wrote:
> On Mon, Jan 25, 2016 at 05:45:30PM -0600, Ben Nemec wrote:
>> On 01/25/2016 03:56 PM, Steven Hardy wrote:
>>> On Fri, Jan 22, 2016 at 11:24:20AM -0600, Ben Nemec wrote:
>>>> So I haven't weighed in on this yet, in part because I was on vacation
>>>> when it was first proposed and missed a lot of the initial discussion,
>>>> and also because I wanted to take some time to order my thoughts on it.
>>>>  Also because my initial reaction...was not conducive to calm and
>>>> rational discussion. ;-)
>>>>
>>>> The tldr is that I don't like it.  To explain why, I'm going to make a
>>>> list (everyone loves lists, right? Top $NUMBER reasons we should stop
>>>> expecting other people to write our API for us):
>>>>
>>>> 1) We've been down this road before.  Except last time it was with Heat.
>>>>  I'm being somewhat tongue-in-cheek here, but expecting a general
>>>> service to provide us a user-friendly API for our specific use case just
>>>> doesn't make sense to me.
>>>
>>> Actually, we've been down this road before with Tuskar, and discovered that
>>> designing and maintaining a bespoke API for TripleO is really hard.
>>
>> My takeaway from Tuskar was that designing an API that none of the
>> developers on the project use is doomed to fail.  Tuskar also suffered
>> from a lack of some features in Heat that the new API is explicitly
>> depending on in an attempt to avoid many of the problems Tuskar had.
>>
>> Problem #1 is still developer apathy IMHO though.
> 
> I think the main issue is developer capacity - we're a small community and
> I for one am worried about the effort involved with building and
> maintaining a bespoke API - thus this whole discussion is essentially about
> finding a quicker and easier way to meet the needs of those needing an API.
> 
> In terms of apathy, I think as a developer I don't need an abstraction
> between me, my templates and heat.  Some advanced operators will feel
> likewise, others won't.  What I would find useful sometimes is a general
> purpose workflow engine, which is where I think the more pluggable mistral
> based solution may have advantages in terms of developer and advanced
> operator uptake.

The API is for end users, not developers.  tripleo-incubator was easily
hackable for developers and power users.  It was unusable for everyone else.

> 
>>> I somewhat agree that heat as an API is insufficient, but that doesn't
>>> necessarily imply you have to have a TripleO specific abstraction, just
>>> that *an* abstraction is required.
>>>
>>>> 2) The TripleO API is not a workflow API.  I also largely missed this
>>>> discussion, but the TripleO API is a _Deployment_ API.  In some cases
>>>> there also happens to be a workflow going on behind the scenes, but
>>>> honestly that's not something I want our users to have to care about.
>>>
>>> How would you differentiate between "deployment" in a generic sense in
>>> contrast to a generic workflow?
>>>
>>> Every deployment I can think of involves a series of steps, involving some
>>> choices and interactions with other services.  That *is* a workflow?
>>
>> Well, I mean if we want to take this to extremes then pretty much every
>> API is a workflow API.  You make a REST call, a "workflow" happens in
>> the service, and you get back a result.
>>
>> Let me turn this around: Would you implement Heat's API on Mistral?  All
>> that happens when I call Heat is that a series of OpenStack calls are
>> made from heat-engine, after all.  Or is that a gross oversimplification
>> of what's happening?  I could argue that the same is true of this
>> discussion. :-)
> 
> As Hugh has mentioned the main thing Heat does is actually manage
> dependencies.  It processes the templates, builds a graph, then walks the
> graph running a "workflow" to create/update/delete/etc each resource.
> 
> I could imagine a future where we interface to some external workflow tool to
> e.g do each resource action (e.g create a nova server, poll until it's active),
> however that's actually a pretty high overhead approach, and it'd probably
> be better to move towards better use of notifications instead (e.g less
> internal workflow)
> 
>>>> 3) It ties us 100% to a given implementation.  If Mistral proves to be a
>>>> poor choice for some reason, or insufficient for a particular use case,
>>>> we have no alternative.  If we have an API and decide to change our
>>>> implementation, nobody has to know or care.  This is kind of the whole
>>>> point of having an API - it shields users from all the nasty
>>>> implementation details under the surface.
>>>
>>> This is a valid argument, but building (and maintining, forever) a bespoke
>>> API is a high cost to pay for this additional degree of abstraction, and
>>> when you think of the target audience, I'm not certain it's entirely
>>> justified (or, honestly, if our community can bear that overhead);
>>>
>>> For example, take other single-purpose "deployment" projects, such as
>>> Sahara, Magnum, perhaps Trove.  These are designed primarily as user-facing
>>> API's, where the services will ultimately be consumed by public and private
>>> cloud customers.
>>>
>>> Contrast with TripleO, where our target audience is, for the most part,
>>> sysadmins who deploy and maintain an openstack deployment over a long
>>> period of time.  There are two main groups here:
>>>
>>> 1. PoC "getting started" folks who need a very simple on-ramp (generalizing
>>> somewhat, the audience for the opinionated deployments driven via UI's)
>>>
>>> 2. Seasoned sysadmins who want plugability, control and flexibility above
>>> all else (and, simplicity and lack of extraneous abstractions)
>>>
>>> A bespoke API potentially has a fairly high value to (1), but a very low or
>>> even negative value to (2).  Which is why this is turning out to be a tough
>>> and polarized discussion, unfortunately.
>>
>> Well, to be honest I'm not sure we can satisfy the second type of user
>> with what we have today anyway.  Our Heat-driven puppet is hardly
>> lightweight or simple, and there are extraneous abstractions all over
>> the place (see also every place that we have a Heat template param that
>> exists solely to get plugged into a puppet hiera file :-).
>>
>> To me, this is mostly an artifact of the original intent of the Heat
>> templates being _the_ abstraction that would then be translated into
>> os-*-config, puppet, or [insert deployment tool of choice] by the
>> templates, and I have to admit I'm not sure how to fix it for these users.
> 
> I think we fix it by giving them a choice.  E.g along the lines of the
> "split stack" approach discussed at summit - allow operators to choose
> either pre-defined roles with known interfaces (parameters), or deploy just
> the infrastructure (servers, networking, maybe storage) then drive
> configuration tooling with a much thinner interface.

I agree, I'm a big fan of the split stack idea to address this problem.
 I don't see that it matters for the API though.  Those power users are
going to use the API only as far as deploying the infrastructure, then
they're going to do their own thing from there.

> 
>> So I guess the question is, how does having an API hurt those power
>> users?  They'll still be able/have to edit Heat templates to deploy
>> additional services.  They'll still have all the usual openstack clients
>> to customize their Ironic or Nova setups.  They're already using an API
>> today, it's just one written entirely in the client.
> 
> There's already a bunch of opaque complexity inside the client and TripleO
> common, adding a very rigid API makes it more opaque, and harder to modify.
> 
>> On the other hand, an API that can help guide a user through the deploy
>> process (You say you want network isolation enabled?  Well here are the
>> parameters you need to configure...) could make a huge difference for
>> the first type of user, as would _any_ API usable by the GUI (people
>> just like pretty GUIs, whether it's actually better than the CLI or not :-).
>>
>> I guess it is somewhat subjective as you say, but to me the API doesn't
>> significantly affect the power user experience, but it would massively
>> improve the newbie experience.  That makes it a win in my book.
> 
> I agree 100% that we need to massively improve the newbie experience - I
> think everybody does.  I also think we also all agree there must be a
> stable, versioned API that a UI/CLI can interact with.
> 
> The question in my mind is, can we address that requirement *and* provide
> something of non-negative value for developers and advanced operators.
> 
> Ryan already commented earlier in this thread (and I agree having seen
> Dan's most recent PoC in action) that it doesn't make a lot of difference
> from a consumer-of-api perspective which choice we make in terms of APi
> impelementation, either approach can help provide the stable API surface
> that is needed.

The problem is I don't see this being true.  We're apparently already
having to change our API and add steps to accommodate the limitations of
the Mistral approach.  This is just trying to duplicate the
functionality we have today, never mind improving on it.

> 
> The main difference is, only one choice provides any flexibility at all wrt
> operator customization (unless we reinvent a similar action plugin mechanism
> inside a TripleO API).

Handing people our code and saying "go ahead and make any changes you
want to this, and don't worry we'll still support your upgrade path when
the next release comes out" doesn't seem like a sustainable model to me.
 The more ad-hoc customization we allow, the worse our encapsulation is
and the more difficult it becomes to guarantee people's cloud will still
work from version to version.  Honestly, we _do_ need a well-defined set
of extension points that people can attach to and that we guarantee will
still be there and functional from release to release.

If we're incapable of doing that as a community then I think we've
failed anyway.  There's too much ongoing cost to exposing all of your
implementation details the way we do today.

> 
>>>> 4) It raises the bar even further for both new deployers and developers.
>>>>  You already need to have a pretty firm grasp of Puppet and Heat
>>>> templates to understand how our stuff works, not to mention a decent
>>>> understanding of quite a number of OpenStack services.
>>>
>>> I'm not really sure if a bespoke WSGI app vs an existing one (mistral)
>>> really makes much difference at all wrt raising the bar.  I see it
>>> primarily as in implementation detail tbh.
>>
>> I guess that depends.  Most people in OpenStack already know at least
>> some Python, and if you've done any work in the other projects there's a
>> fair chance you are familiar with the Python clients.  How many people
>> know Mistral YAML?
> 
> So, I think you're conflating the OpenStack developer community (who,
> mostly, know python), with end-users and Operators, where IME the same is
> often not true.
> 
> Think of more traditional enterprise environments - how many sysadmins on
> the unix team are hardcore python hackers?  Not that many IME (ignoring
> more devops style environments here).
> 
>> Maybe I'm overestimating the Python knowledge in the community, and
>> underestimating the Mistral knowledge, but I would bet we're talking
>> order(s?) of magnitude in terms of the difference.  And I'm not saying
>> learning Mistral is a huge task on its own, but it's one more thing in a
>> project full of one more things.
> 
> It's one more thing, which is already maintained and has an active
> community, vs yet-another-bespoke-special-to-tripleo-thing.  IMHO we have
> *way* too many tripleo specific things already.
> 
> However, lets look at the "python knowledge" thing in a bit more detail.
> 
> Let's say, as an operator I want to wire in a HTTP call to an internal asset
> management system.  The requirement is to log an HTTP call with some
> content every time an overcloud is deployed or updated.  (This sort of
> requirement is *very* common in enterprise environments IME)
> 
> In the mistral case[1], the modification would look something like:
> 
> http_task:
>   action: std.http url='assets.foo.com' <some arguments>

requests.get('assets.foo.com', data={'some':'arguments'})

Either way you have to go look up the exact API to figure out how to
make the call the way you need.

> 
> You'd simply add two lines to your TripleO deployment workflow yaml[2]:
> 
> Now, consider the bespoke API case.  You have to do some or all of the
> following:
> 
> - Find the python code which handles deployment and implements the workflow

-Find the Mistral YAML which handles deployment and implements the workflow.

> - Pull and fork the code base, resolve any differences between the upstream
>   version and whatever pacakged version you're running

-I'm not sure why you would rebase on upstream.

> - Figure out how to either hack in your HTTP calls via a python library, or
>   build a new plugin mechanism to enable out-of-tree deployment hooks

-As noted above, it's not exactly difficult.  If you know neither
Mistral nor Python, you're learning a new API either way.  Except one is
way more broadly useful.

> - Figure out a bunch of complex stuff to write unit tests, battle for
>   weeks/months to get your code accepted upstream (or, maintain the fork
>   forever and deal with rebasing, packaging, and the fact that your entire
>   API is no longer supported by your vendor because you hacked on it)

-Or we could design a decent API where users don't have to hack up our
"code" to do what they need.  If this is so common, why isn't it just
part of the deployment API?

> 
> Which of these is most accessible to a traditional non-python-ninja
> sysadmin?
> 
> [1] http://docs.openstack.org/developer/mistral/dsl/dsl_v2.html#std-http
> [2] https://github.com/dprince/tripleo-common/blob/mistral/workflows/overcloud_deploy.yaml
>