On 05/02/2011 12:38, Guillaume Nodet wrote:
I think we may want to stick too much to semantic versioning.
Is there a 'not' missing?
Don't
get me wrong, I agree with the compatibility rules semantic and we
should support them: i.e. if there is an incompatible change, the
major version has to be different and if there is a compatible change,
the minor version has to be different.
However, I think it does not imply we need to be strict about *when*
we change the version of the package. If we align the package version
changes with the bundle version changes (that does not necessarily
imply they have to be the same), i.e. when we change from aries 0.3 to
aries 0.4, the packages are bumped too from a minor version, this
leaves some place to do branch fixes using minor versions on the
package.
I think we face two different development model basically:
* use a more traditional development model using a tree of versions
(either at the component level or for the whole tree). We can support
branch fixes and micro releases in a tree. We need to somewhat align
the package version changes to the bundle version change. This means
we may end up with multiple identical packages with the same version
number (though I suppose real api and library packages are somewhat
different, as pure API packages that do not contain any code could
have their own versioning not tied to this). We can release / branch
per component or from the root (still to be decided, but completely
orthogonal to the problem)
I think we need to be able to articulate some clear 'rules' on versions.
* use the model that felix / sling follows and have a versioning
scheme per bundle with no way to maintain bug fix branches. Each
bundle has to be released independantly with its own release notes and
all. I think we'd have to flatten our svn tree as there's no real
need to have a tree of modules (we can use maven profiles if we want
to build only a given component). The consequence is that we also
need to maintain documentation about compatiblity, as users see the
bundle versions, not really the package versions and we need to have a
way to tell them which bundles are supposed to work together (i think
we have more than 100 bundles in the whole tree I think, so not sure
what kind of documentation could give an overview of that ;-) ).
So, just to consider blueprint, which currently has 12 bundles, we would
- release each bundle separately, with its own vote. Its own JIRA
component, its own release notes, its own release announce etc.
- say we have blueprint-something at version x.y.z-SNAPSHOT, and we
want to release version x.y.z. In trunk there is also
blueprint-anotherthing at version a.b.c-SNAPSHOT. If we release
blueprint-something x.y.z, what do we say about what it has been
built/tested with? Or maybe we have to release blueprint-anotherthing
a.b.c at the same time?
It seems any in-between solution has some big problems. Currently,
we're more on the first solution and I think it's way easier for us
and for our users.
It's easier for us. But what we do at the moment is definitely wrong wrt
to package versions. Is the implication of what you say above that you
believe we should not use semantic versioning (strictly) for packages?
Or do you think we can achieve this with some modifications to the
current process? That's what I was trying to explore. Independent bundle
releases does not seem practical (for us) unless we opted to combine
everything in a single uber bundle (and I think there were arguments
against that).
Zoe
On Sat, Feb 5, 2011 at 11:03, Felix Meschberger<[email protected]> wrote:
Hi,
Am Samstag, den 05.02.2011, 00:43 +0100 schrieb Guillaume Nodet:
In fact I want it to be easy, the easier the better. If you change the code the
version should increment.
So why even bother with having to manually change the version ? I
think it should be possible to have the maven-bundle-plugin increment
the version depending on the kind of changes by comparing the package
signatures and make sure it follows the semantic versioning. Given it
has already been done (see
https://www.assembla.com/wiki/show/obcc/OSGi_Version_Generator) I
think we could add that to the maven bundle plugin.
+1 !
(Not sure, whether the Bundle Plugin does not get overloaded, though)
However, there's something which is worrying me about the semantic
versioning. I don't think it can cope with maintenance branches. The
process of incrementing a package version works well in a single line
of releases, but not in a tree, so can't ever release a package which
doesn't contain all the previous changes. Or rather the process works
for a given package, but the problem is that our bundles do not only
contain a single package. Let's take a concrete example.
I have a bundle version 3.0 which has two packages foo / 1.0 + bar / 1.0.
In a future release 3.1 of that bundle, i add one functionality to the
foo package and a minor modification to the bar package, so I release
this with foo / 1.1 + bar / 1.0.1.
Some time later, I find a bug in the bar package which I'd like to fix
for both minor versions of my bundle. If I do so, I'd end up with a
bundle 3.0.1 with foo / 1.0 + bar / 1.0.1 and a bundle 3.1.1 with foo
/ 1.1 + bar / 1.0.2. That's not really possible because the two bar /
1.0.1 package would be different.
Possible solutions:
* backport into 3.0 branch the change that modification that caused
the bump from 1.0 to 1.0.1 (when releasing 3.1). However, this may
not be a big fix, maybe a small improvement that I don't want to
backport, so I don't think this solution is a good idea
* never release a bundle which exports multiple packages: that sucks too
* don't do maintenance release: i don't think we want that
* consider that any modification you may not want to backport in a
maintenance branch later should lead to bump the minor version of a
package, even if the signature of the package doesn't really change
I think the last one is the only one applicable. Thoughts ?
Well branches pose just more than this single problem, which is why I
generally try to avoid release branches like the plague ...
In this concrete example, you might probably have to either backport
everything from the bar/1.0.1 package or employ qualifier increments.
At the end of the day, this should not prevent the project from adopting
semantic versioning, because the benefits tremendously outweigh these
minor costs.
Regards
Felix
On the other hand, the problem is always the same: you have to update
information in a secondary location -- regardless of whether this is the
packaginfo or the pom.xml file.
Right, and packageinfo is right next to the classes you just updated, closer =
better IMHO.
The advantage of doing it in the pom.xml file IMHO is that you have a
complete overview of your exports incl. their versions. YMMV.
The downside in our case is you have to update the bundle and the uber bundle,
bigger chance of getting out of sync, which would be very very bad.
also you need to sync the version correctly
between the bundles
the bundle plugin takes care of this (fortunately) -- assuming you mean
the "Import-Package" versioning.
I'm trying to address export bundle. I'm happy with the import package stuff.
and the uber bundles.
bnd supports the packageinfo files (and also annotations in
package-info.java), but those are not currently picked up and used in
our build. I raise FELIX-2819 and a workaround has been suggested,
which I managed to get working.
The fix would be to add the following to the default-pom and get the
modules to use the updated parent:
<resource>
<directory>${project.build.sourceDirectory}</directory>
<includes>
<include>**/packageinfo</include>
</includes>
</resource>
Unfortunately, you will still have the regular resources in the
src/main/resources tree. So you have to explicitly list this to in the
<resources> element of the parent POM to not miss these...
Right, this is already in our parent pom, so I'm proposing adding this in
addition to the other resources statements already there.
In Sling we currently maintain the exported package version in the POMs.
This works fine but is also kind of suboptimal.
I think the most important thing is to make it consistent: Do it either
way, but stick to.
I agree. I think we should use packageinfo though :)
Regards
Felix
Thoughts?
Alasdair
