[openstack-dev] [Murano] Versioning of Murano packages and MuranoPL classes
Georgy Okrokvertskhov
gokrokvertskhov at mirantis.com
Fri Jul 10 14:56:43 UTC 2015
Hi Alex,
Thank you for the great summary.
I have a concern about item #8. Can we add an option to Murano to use
previous storage engine rather then Glance v3? We need to make sure that v3
API in Glance is set by default before we do a hard dependency on it in
Murano.
Thanks
Gosha
On Fri, Jul 10, 2015 at 4:53 AM, Alexander Tivelkov <ativelkov at mirantis.com>
wrote:
> Hi folks,
>
> Ability to manage multiple versions of application packages and their
> dependencies was always an important item in Murano roadmap, however we
> still don't have a clear spec for this feature.
> Yesterday we hosted a small design session to come up with a plan on what
> can be done in Liberty timeframe to have proper versioning for MuranoPL
> classes and packages. Stan Lagun, Kirill Zaitsev and myself participated
> offline, some other muranoers joined remotely. Thanks to everybody who
> joined us.
>
> TL;DR: it turns out that now we have a clear plan which will allow us to
> achieve proper versioning of the packages and classes, and we'll try to
> implement the most important parts of it in Liberty.
>
> Here is the detailed outcome of the session:
>
>
> 1. We'll use the standard Semantic Versioning format
> ('Major.Minor.Patch[-dev-build.label[+metadata.label]]') to version our
> packages: changes which break backwards-compatibility should increment the
> major segment, non-breaking new features increment the minor segment and
> all non-breaking bugfixes increment the patch segment. The developers
> should be carefull with the "new features" part: if you add a new method to
> a class, it may be considered a breaking change if the existing subclasses
> of this class have the same method already declared. We still assume that
> such changes should lead to increment of 'minor' segment, however it is up
> to best judgement of developers in particular case: if the risk of such
> method override is really high it may worth to increment the 'major'
> segment. Proper guideline on the versioning rules will be published closer
> to L release.
>
> 2. A new field 'Version' is introduced into package manifest file
> which should define package version in complete semver format. The field
> itself is optional (so existing apps are not broken), if it is not
> specified the package is assumed to have version '0.0.0'
>
> 3. The existing 'Require' block of Application manifest will be used
> to specify the package dependencies. Currently it is a yaml-based
> dictionary, with the keys set to fully-qualified names of the dependency
> packages and the values set to the version of those dependencies. Currently
> this block is used only for integration with apps stored at
> apps.openstack.org. It is suggested to use this block in the
> deployment process as well, and extend its semantics.
> The version of the dependency specified there should also follow the
> semver notation, however it may be specified in the shortened format, i.e.
> without specifying the 'patch' or 'patch' and 'minior' components. In this
> case the dependency will be specified as a range of allowed versions. For
> example, a dependency version 1.2 will mean a (1.2.0 >= version < 1.3)
> range.
> If the version of a dependency is not specified (like in this existing
> app - [1]) then we assume the version "0" - i.e. the last available
> pre-release version of a package.
>
> 4. Murano core library is also a package which has its own version.
> The current one is assumed to have a version 0.1.0, the one which is going
> to be released in L will be probably called 0.2.0. The lib is still quickly
> evolving, so we are not releasing a 1.0.0 until we are sure that we are not
> going to have any breaking changes anytime soon.
> As with any other package it will be possible to have several versions
> of the Core Library installed in Murano at the same moment of time.
>
> 5. There is no mandatory need to add the the dependency on the core
> library to the "Requires" block of each application, as it is added there
> implicitly. However, this implicit dependency will have a version "0" -
> i.e. will reference the latest pre-release version of the Core Library
> available. So it is still better to pin the core library requirement to a
> particular version to make sure that your app does not break if we
> introduce any breaking change into the core lib.
>
> 6. All classes defined in a package are assumed to have a version
> identical to the version of the package.
>
> 7. Murano Extension Plugins (i.e python packages which declare
> setuptools-entrypoints in 'io.murano.extensions' namespace) also will have
> similar versioning semantics: they will have a fully qualified name
> (defined by the setuptools' package name) and a version (also defined by
> setuptools), an will get an ability to specify their own dependencies if
> needed. From the class-loader perspective the MuranoPL classes defined in
> the plugins are no difference from the classes defined in a regular package.
>
> 8. We are going to store murano packages as Glance V3 Artifacts
> Repository, naturally mapping package's FQN and version to artifact's name
> and version.
> The package dependencies will be stored in Glance as cross-artifact
> dynamic dependencies (i.e. dependencies not on a particular artifact but on
> the last artifact matching the given name and the version range query) as
> soon as that feature is implemented in Glance (currently only static
> dependencies are implemented there). Until that, the dependencies will be
> stored as a regular list of strings, and the Murano engine will process it
> and query Glance to fetch the packages.
>
> 9. In L cycle we are not going to show multiple versions of the same
> app in Murano dashboard: only the last one will be shown if the multiple
> versions are present. This is to minimize the changes at Dashboard side: in
> future releases we'll add the ability to select the proper version.
> The generation of the object model by dynamic UI also remains intact.
>
> 10. However, the structure of the object model isself gets changed: in
> the "?" block of each object two new fields appear: "package" and
> "version", which correspond to the FQN and the version of the package which
> contain the class of the given object. UI leaves these fields as Nones when
> it generates the OM, and the engine computes them in a regular way: queries
> the package repository for the most recent version of a package which
> contains the class with a given name, and saves information about its name
> and version. This values get persisted in an Object Model when it gets
> serialized after the deployment. As a result, the versions of the
> components are fixed once the environment is deployed, so if some packages
> get updated afterwards, the existing components remain pinned to their
> initial version. As a result, the environment may get several components of
> the same type but different versions.
>
> 11. When the Object Model is validated after the deserialization, the
> behavior of "$.class()" contract is changed. During its validation the
> value passed to the appropriate property or argument should be of a type
> which is declared either in a current package (or in the another version of
> the current package, given that the major component of the versions is the
> same) or in one of the packages satisfying the requirements of the current
> one. I.e. it becomes impossible to reference any class from the
> unreferenced package.
>
> 12. When inheriting some other class using the 'Extends' attribute,
> the ancestor class should be defined either in the current package or in
> one of the packages satisfying the requirements of the current one.
>
> 13. (creepy advanced stuff) It may turn out that in case of the
> multiple inheritance a single class will attempt to inherit from two
> different versions of a same class. An exception should be thrown in this
> case, unless there is a possibility to find a version of this class which
> satisfies all parties.
>
> *For example: classA inherits classB from packageX and classC from
> packageY. Both classB and classC inherit from classD from packageZ, however
> packageX depends on the version 1.2.0 of packageZ, while packageY depends
> on the version 1.3.0. This leads to a situation when classA transitively
> inherits classD of both versions 1.2 and 1.3. So, an exception will be
> thrown. However, if packageY's dependency would be just "1" (which means
> any of the 1.x.x family) the conflict would be resolved and a 1.2 would be
> used as it satisfies both inheritance chains.*
>
>
>
> So, all the above cover most of our present needs for MuranoPL package and
> class versioning.
> Also, we already have a way which allows us to properly version the format
> of MuranoPL language (a "Format" key in application manifests) and
> UI-definition files ("Version" key in that files). This basically allows us
> to target the packages for a minimum version of Murano / Murano Dashboard.
>
>
> I hope this rather lengthly email is useful. Stan Lagun has taken an
> action item to frame all the above into a more formal spec.
>
>
>
> [1]
> https://github.com/openstack/murano-apps/blob/master/MySQL/package/manifest.yaml#L26
>
>
> --
> Regards,
> Alexander Tivelkov
>
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
--
Georgy Okrokvertskhov
Architect,
OpenStack Platform Products,
Mirantis
http://www.mirantis.com
Tel. +1 650 963 9828
Mob. +1 650 996 3284
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20150710/d881414e/attachment.html>
More information about the OpenStack-dev
mailing list