On 06/20/2015 02:14 AM, Devananda van der Veen wrote:
Almost all of our discussions so far on this topic have left something
out, which Monty pointed out to me last week. I'm following up now
because E_TRAVEL...
tldr;
What we're versioning here are API's, not packages. It's not a question
of numbering and dependency ordering, but of communicating
support[ed|able] interfaces between running services. Libtool is more
relevant than semver.
http://www.gnu.org/software/libtool/manual/html_node/Updating-version-info.html
Right now we lack a means to express that the API response is
"compatible-with" a particular previously-released version of the API.
If we had that, instead of "current-version", I believe we would have
much happier users (and a happier Dmitry and jroll).
Long version...
Every HTTP response from Ironic today includes three headers: min, max,
and version. The service can present an older API version, as long as it
is greater-than-or-equal-to the minimum supported version, even if that
version is incompatible with the maximum supported version. It does
this by rewriting responses to match what was expected in the requested
(older) version.
When the newer version is identical *for all interfaces present* in the
older version, this can be called compatible. Dmitry's point is that we
don't need to hide newer interfaces from users who request an older API
version, because the client won't know or care about things that weren't
in the version it requested.
+1
However, we *do* need to signal their presence, and we don't have a good
means for that right now. We also need to signal to the client that the
response given is "compatible with" a certain "age" of API, even if it's
not exactly the same. And we don't have any means for that, either.
Time for an example....
Let's say that an incompatible change was made in v1.3. Let's also say
that a change was made in v1.5 that added a new endpoint. Today, this is
what the response headers would look like when calling a server running
v1.5.
a) client requests v1.2, receives headers (min: 1.1, max: 1.5, current:
1.2) b) client requests v1.4, receives headers (min: 1.1, max: 1.5,
current 1.4) c) client requests v1.5, receives headers (min: 1.1, max:
1.5, current: 1.5)
What we have implemented today is that in case (b), the service will
*hide* any changes that we introduced in v1.5. But those changes did not
affect any functionality of the v1.4 API, so Dmitry objects to this. So
do I.
The issue at hand is the response in case (b) ... though after spending
the last several months working on api versioning, I actually think all
of those are poor responses.
What I believe we should have:
a) client requests v1.2, receives headers (min: 1.1, max: 1.5,
compatible-with: 1.1)
b) client requests v1.4, receives headers (min: 1.1, max: 1.5,
compatible-with: 1.3)
b) client requests v1.5, receives headers (min: 1.1, max: 1.5,
compatible-with: 1.3)
That sounds really like something people actually want from us.
Just one clarifying question: shouldn't b) and c) be compatible-with:
1.4, as 1.4 was a breaking change?
Yes -- (b) and (c) are identical responses.
Discuss.
-Devananda
On Tue, Jun 16, 2015 at 7:13 AM Dmitry Tantsur <[email protected]
<mailto:[email protected]>> wrote:
On 06/16/2015 03:47 PM, Jim Rollenhagen wrote:
> On Tue, Jun 16, 2015 at 08:56:37AM +0200, Dmitry Tantsur wrote:
>> On 06/04/2015 08:58 AM, Xu, Hejie wrote:
>>> Hi, guys,
>>> I’m working on adding Microversion into the API-WG’s guideline
which
>>> make sure we have consistent Microversion behavior in the API
for user.
>>> The Nova and Ironic already have Microversion implementation,
and as I
>>> know Magnum _https://review.openstack.org/#/c/184975/_ is going to
>>> implement Microversion also.
>>> Hope all the projects which support( or plan to) Microversion
can join
>>> the review of guideline.
>>> The Mircoversion specification(this almost copy from nova-specs):
>>> _https://review.openstack.org/#/c/187112_
>>> And another guideline for when we should bump Mircoversion
>>> _https://review.openstack.org/#/c/187896/_
>>> As I know, there already have a little different between Nova and
>>> Ironic’s implementation. Ironic return min/max version when the
requested
>>> version doesn’t support in server by http-headers. There isn’t such
>>> thing in nova. But that is something for version negotiation we
need for
>>> nova also.
>>> Sean have pointed out we should use response body instead of http
>>> headers, the body can includes error message. Really hope
ironic team
>>> can take a
>>> look at if you guys have compelling reason for using http headers.
>>> And if we think return body instead of http headers, we
probably need
>>> think about back-compatible also. Because Microversion itself isn’t
>>> versioned.
>>> So I think we should keep those header for a while, does make
sense?
>>> Hope we have good guideline for Microversion, because we only
can change
>>> Mircoversion itself by back-compatible way.
>>> Thanks
>>> Alex Xu
>>
>> Hi all!
>>
>> I'd like to try put in feedback based on living with
microversions in Kilo
>> release of Ironic.
>
> And here's my take, based on my experiences. Keep in mind I'm a core
> reviewer, a developer, and an operator of Ironic.
Thanks Jim, much appreciated!
>
> From an ops perspective, our team has built our fair share of
tooling to
> help us run Ironic. Some of it uses the REST API via python or
node.js,
> and of course we all use the CLI client often.
>
> We also continuously deploy Ironic, for full transparency. My
experience
> is not with how this works every 6 months, but in the day-to-day.
>
>>
>> First of all, after talking to folks off-list, I realized that
we all, and
>> the spec itself, confuse 3 aspects of microversion usage:
>>
>> 1. protecting from breaking changes.
>> This is clearly a big win from user's point of view, and it
allowed us to
>> conduct painful change with renaming an important node state in
our state
>> machine. It will allows us even worse change this cycle: change
of the
>> default state.
>>
>
> +1. Good stuff. My tooling doesn't break when I upgrade. Yay.
>
>> 2. API discoverability.
>> While I believe that there maybe be better implementation of
this idea, I
>> think I got it now. People want services to report API versions they
>> support. People want to be able to request a specific version,
and fail
>> early if it is not present. Also +1 from me.
>>
>
> I don't tend to personally do this. I usually am aware of what
version
> of Ironic I'm running against. However I see how this could be useful
> for other folks.
>
> I do, however, use the versions to say, "Oh, I can now request
1.5 which
> has logical names! That's useful, let's set those to the names in our
> CMDB." Now my tooling that interacts with the CMDB and Ironic can
look
> at the version and decide to use node.name <http://node.name>
instead of the old hack we
> used to use.
>
>> 3. hiding new features from older clients
>> This is not directly stated by the spec, but many people imply
it, and Nova
>> and Ironic did it in Kilo. I want us to be clear: it is not the
same as #2.
>> You can report versions, but still allow new features to be used.
>>
>
> This is still totally useful. If you know what version you are
running
> against, you know exactly what features are available.
"You know" is about #2 - that's where confusion is :)
so if you know, that moving to inspection state is disallowed for your
tooling (but not for the whole system!), what does it give you?
>
> I think the disconnect here is that we don't expect users
(whether those
> are people or computers) to explicitly request a version. We need to
> message better that if you are using Ironic or building a tool
against
> Ironic's API, you should be pinning the version. We also need to take
> this comment block[0] and put it in our docs, so users know what each
> version does.
>
> Knowing that I get feature X when I upgrade to version Y is useful.
>
>> It is this particular thing that gets -2 from me, after I've
seen how it
>> worked in practice, and that's why.
>>
>> First of all, I don't believe anyone needs it. Seriously, I
can't imagine a
>> user asking "please prevent me from using non-breaking changes".
And attempt
>> to implement it was IMO a big failure for the following reasons:
>>
>> a) It's hard to do. Even we, the core team, got confused, and
for non-core
>> people it took several iteration to do right. It's a big load on
both
>> developers and reviewers.
>>
>
> I do agree with this. It's been painful. However, I think we're
mostly
> past that pain at this point. Does this patch[1] look like developer
> pain?
It's not that painful to write. Now. When we have 10-20 version, it
probably will :)
anyway, it's hard to explain newcomers how to do it, and it's hard to
review the result. we failed at it, e.g. with error codes.
>
>> b) It's incomplete (at least for Ironic). We have several
API-exposed things
>> that are just impossible to hide. Good example are node states:
if node is
>> in a new state, we can't but expose it to older tooling. Our
free-form JSON
>> fields properties, instance_info, driver_info and
driver_internal_info are
>> examples as well. It's useless to speak about API contract,
while we have
>> those.
>>
>
> I somewhat agree here.
>
> With node states, there are cases where we were able to hide it
> (NOSTATE -> AVAILABLE), and cases where we were not (adding
MANAGEABLE).
> However, this list of states is (AIUI) not part of the API contract;
> rather the verbs available to move between states are.
What's the point in contract, if there are things not covered by it that
drastically change the system behavior?
>
> As far as JSON fields, we've never had a contract around what
keys are
> available. Only the semantics of working with those fields, and which
> fields exist.
ditto as above: you can request new features by modifying driver_info.
>
>> c) It gives additional push back to making (required) breaking
changes. We
>> already got suggestions to have ONE MORE feature gating for breaking
>> changes. Reason: people will need to increase microversions to
get features,
>> and your breaking change will prevent it.
>>
>
> This is just silly. If 1.10 breaks a user, and the user wants 1.11,
> they'll need to fix that breakage.
++ but not everyone agreed on the summit, when I was talking about
ENROLL state
>
>> d) It requires a hard compromise on the CLI tool. You either
default it to
>> 1.0 forever, and force all the people to get used to figuring
out version
>> numbers and using `ironic --ironic-api-version x.y` every time
(terrible
>> user experience), or you default it to some known good version,
bumping it
>> from time to time. This, in turn, has 2 more serious problems:
>>
>
> I disagree that pinning a version all the time is a terrible
experience.
> We already require a number of options for authentication
(OS_USERNAME,
> OS_PASSWORD, etc etc). How many folks do you think type these in
every
> time? Solution is simple: add IRONIC_API_VERSION to whatever
exports the
> other environment variables.
It's not that bad, especially if devstack/tripleo will provide some
reasonable default for you.
I remember, however, Devananda didn't like the idea.
And it definitely makes a quick start guide a bit harder to follow. I
already imagine how many people will forget about this pinning (either
to do it, or do update when they need new features).
>
> The version depends on the environment you are running against -
why not
> treat it as such?
>
>> d.1) you start to break people \o/ that's not a theoretical
concern: our
>> downstream tooling did get broken by updating to newer
ironicclient from git
>>
>
> As I said before, we need to encourage folks to pin client
versions if
> they don't want to break. I'm probably alone here, but I would even
> propose making the version *required*. Force people to think
about what
> they are doing. If folks are okay with being broken, they can pass
> "latest".
Could be a good default for devstack btw
>
>> d.2) you require complex version negotiations on the client
side. Otherwise
>> imaging CLI tool defaulting to 1.6 will issue `node-create` to
Ironic
>> supporting only 1.5. Guess what? It will fail despite
node-create being very
>> old feature. Again, that's not something theoretical: that's how
we broke
>> TripleO CI.
>>
>
> Again, pin it.
>
>> e) Every microversion should be fully tested separately. Which
ended up in
>> Ironic having 4 versions 1.2-1.5 that were never ever gate
tested. Even
>> worse, initially, our gate tested only the oldest version 1.1,
but we solved
>> it (though it took time to realize). The only good thing here is
that these
>> versions 1.2-1.5 were probably never used by anyone.
>>
>
> Hi. I've used some of these. :)
You didn't tell me last time we talked :) note, that you didn't use
them, unless you explicitly requested, because IIRC we never defaulted
our client to any of these. So for most people, even deploying from
master, it was 1.1 -> 1.6.
>
> // jim
>
> [0]
https://github.com/openstack/ironic/blob/master/ironic/api/controllers/v1/__init__.py#L59-63
> [1]
https://review.openstack.org/#/c/188873/1/ironic/api/controllers/v1/node.py
>>
>> To sum this long post up, I'm seeing that hiding new features
based on
>> microversions brings much more problems, than it solves (I'm not
aware of
>> the latter at all). I'm very opposed to continuing doing it in
Ironic, and
>> I'm going to propose patch stopping gating Kilo changes
(non-breaking
>> obviously).
>>
>> Hope that helps,
>> Dmitry
>>
>>>
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: [email protected]?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
__________________________________________________________________________
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: [email protected]?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
