I think we are saying the same thing -- we evolve project model used
during the build but deploy both the new and backwards compatible models.
One quick note on representing dependencies as provided/required
capabilities. Although I like this idea in general, I believe it will
require completely new repository layout to be able to efficiently
find capability providers. Single repository-wide metadata index file,
the approach implemented in P2 for example, won't scale for repositories
of the size of Central, so most likely the new repository layout will
require active server-side component to assist dependency resolution.
--
Regards,
Igor
On 11/24/2013, 4:25, Stephen Connolly wrote:
On Sunday, 24 November 2013, Igor Fedorenko wrote:
On 11/23/2013, 23:08, Jason van Zyl wrote:
On Nov 23, 2013, at 5:44 PM, Stephen Connolly
<stephen.alan.conno...@gmail.com> wrote:
Before I forget, here are some of my thoughts on moving towards
Model Version 5.0.0
The pom that we build with need not be the pom that gets
deployed... thus the pom that is built with need not be the same
format as the pom that gets deployed.
Can you explain why you think this is useful? To me all the
information that is carried with the POM after deployment is
primarily for the consumption of tools, and there are a lot of tools
that expect more than the dependency information. Removing all other
elements in the POM is equivalent to being massively backward
incompatible for an API. And if the subsequent consumption after
deployment is primarily by programs, then why does it matter what
gets deployed. I don't really see much benefit, but will create all
sorts of technical problems where we need multiple readers and all
that entails and the massive number of problems this will cause
people who have created tooling, especially IDE integration. >
The way I see it, what is deployed describes how the artifact needs to
be consumed. This is artifact's "public API", if you will, it will be
consumed by wide range of tools that resolve dependencies from Maven
repositories and descriptor format should be very stable. Mostly likely
we have no choice but use (a subset of) the current 4.0.0 model version.
I would be very sad if we are limited to a subset.
There are some critical concepts that in my view are missing from pom files.
Number one on my hit list is a <provides> concept.
Where you declare that an artifact *provides* the same api as another GAV.
Technically you'd need to be able to specify this both at the root of a pom
and also flag specific dependencies (because the api they provide was not
specified when that pom was deployed)
Thus the Geronimo specs poms could all <provides> the corresponding JavaEE
specs and excludes issues or other hacks would no longer be required.
Look at the issues you will have if you use the excludes wildcards in your
pom... Namely *anyone* who uses your artifact as a dependency will need to
be using Maven 3 or newer... does ivy read those wildcards correctly? Does
sbt? Does Buildr?
They are a tempting siren... And from another PoV they will force others to
follow... *but* if we are forcing them to follow should we not pick a nicer
format to follow... Not one consisting of many layers of hacks?
The modelVersion 4.0.0 pom is deployed to the repo (in my scheme) so that
legacy clients can still make some sense... If a modelVersion 5.0.0 feature
cannot be mapped down to 4.0.0... Well we try our best and that's what you
get... We should make it sure that people stuck with older clients can read
something semi-sensible and then layer their hacks as normal to get the
behaviour they need.
Thus <provides> (which is not a scope but a GAV) can be modelled by having
the modelVersion 4.0.0 pom list the collapsed dependencies with the
appropriate <excludes> added (without wildcards)
Other concepts cannot be mapped, so they get dropped.
How the artifact is produced, on the other hand, is artifact's
implementation detail. It is perfectly reasonable for a project to
require minimal version of Maven, for example. Or use completely
different format, not related to pom at all.
Exactly... The pom used to build can be written in JSON or whatever domain
specific DSL you want... We deploy a modelVersion 5.0.0 pom as XML because
it will be read my machines.
Now for the reason I think deploying a pom as xml may be a good thing...
XSLT
Suppose we specify a XSLT GAV that will down-map the pom to a modelVersion
5.0.0 pom... Now we can actually deploy a modelVersion 7.3.5 pom to the one
GAVCT and a modelVersion 5.0.0 client reads is, sees it is a modelVersion
that is not understood, sees the GAV of the XSLT, pulls it down and
transforms the model into the version it can parse
Will it be able to parse all the info in the original pom? Nope... It's an
older client... Older clients should not expect to grok all the subtleties
of newer poms... But it should get the general shape
In all of the above, <packaging>pom</packaging> is special... We just
deploy that as is in whatever format (JSON/DSL/XML/groovy/etc) as the
-build.pom
So 4.0.0 => .pom
5.0.0 onward (XSLT down versioning) => -dep.pom
And as a parent =< -build.pom
Modern clients can ask for the -dep.pom first... And fall back to the .pom
It's not perfect, but it should not be the hell of 3.0.0->4.0.0 the fear of
which has prevented forward progress since
By separating "consumption" and "production" metadata formats, we'll be
able to evolve production format more aggressively. For example, it
would be nice to have Tycho-specific configuration markup inside <build>
section. This is not currently possible because all poms must be
compatible with the same model.
--
Regards,
Igor
Only with <packaging>pom</packaging> do we actually need things like the
<plugins> section in the deployed pom, because it is a reality that for
noo-pom packaging we just want the transitive dependencies.
Now there is the <extensions> issue where you might be registering a
different file type that has different rules with respect to the
classpath... but I am unsure if we actually consider those when
evaluating
the dependency tree... and in any case, once we accept that the deployed
pom is not the same as the pom used to build (for non-pom packaging at
least) we can transform that dependency tree using the exact rules that
have to be known at build time thus closing the extensions issue.
For projects with <packaging>pom</packaging> in fact we are only
deploying
smal files so perhaps we can deploy two pom files... the one that exposes
the standard dependency stuff and then a second one that is used for
build
inheritance.
My vision is thus that we deploy between 2 and three pom files for every
project.
For jar/war/ear/... we deploy:
* a modelVersion 4.0.0 pom as .pom (only lists dependencies)
* a modelVersion 5.0.0 pom as -v5.pom (only lists dependencies but allows
for new scopes)
For pom we deploy
* a modelVersion 4.0.0 pom as .pom (only lists dependencies)
* a modelVersion 5.0.0 pom as -v5.pom (only lists dependencies but allows
for new scopes)
* the pom itself
When building a pom, your parent pom must be of a modelVersion <= your
modelVersion.
Thanks,
Jason
----------------------------------------------------------
Jason van Zyl
Founder, Apache Maven
http://twitter.com/jvanzyl
---------------------------------------------------------
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@maven.apache.org
For additional commands, e-mail: dev-h...@maven.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@maven.apache.org
For additional commands, e-mail: dev-h...@maven.apache.org
