Following up on my previous response, and in a more realistic way than
switching to OSGi, I wonder if we should let maven 4 plugins do their own
DI injection and only care about the mojos, i.e. not rely on sisu/plexus
to
create and inject the mojo. Mojos using the v4 api do not need to be
injected with other components from maven, they should all come from the
api and should be retrieved using Session.getService(xx), or simply using
session's shortcut methods. The injection of Project and Session is not
controlled by sisu. For the ComponentConfigurator, we could change the
mojo
descriptor to have the full configuration class name for mojos that
require
custom configuration injection, the plugin manager being in charge of
instantiating this class and using it as a ComponentConfigurator (which is
not part yet of the api btw).
Complex plugins which rely on plexus/sisu to do some custom injection
would
have to be changed to do their own DI, maybe using a simple Guice setup.
If that sounds like too big a change for those plugins, we may be able to
add a flag on the mojo descriptor so that those v4-enabled mojos would
trigger a DI injection on behalf of the plugin. But if we have to
implement something like that, I would go for a plain CDI-like api, either
using guice or another DI library supporting the javax.inject package, or
rather the jakarta.inject package, as it would be nice to switch to the
jakarta version at the same time. Or maybe even plexus if we really need
to, but with a limited scope, i.e. no visibility on the maven components,
so that plugins are better decoupled from maven-core.
Thoughts ?
Le jeu. 17 nov. 2022 à 17:48, Guillaume Nodet <gno...@apache.org> a
écrit :
I do agree that debugging the provisioning side is *very* complicated
when
there's a problem.
I'd be happy to get rid of sisu/plexus and use a more simple DI
framework,
at least for simple plugins.
However, I definitely don't think pushing OSGi to plugins would be a
good
idea : the cost and burden on plugin developers would outweigh the
benefits.
For extensions, and for maven itself, that is a different story though.
Maven and extensions could definitely benefit from OSGi, but this would
be
a complete breakage and it would be hard to keep some level of
compatibility.
Le jeu. 17 nov. 2022 à 17:00, Christoph Läubrich <m...@laeubi-soft.de>
a
écrit :
> Guess classrealm is fine for maven, it does not bring much issues
As long as it works, maybe, maybe even if you write a simple maven
plugin, but for any more complex it is just a complete mess.
Last time I asked on the mailing list how to debug that stuff ...
complete silence ...
Today I tried to refactor the name of one module of a more complex
maven-plugin (with core extension), now I end up with
org.apache.maven.InternalErrorException: Internal error:
com.google.inject.ProvisionException: Unable to provision, see the
following errors:
1) No implementation for org.eclipse.aether.RepositorySystem was bound.
A whole bunch of stack trace but not a little info why the ***** it is
not happy. Now I need to add random "exportedArtifact" /
"exportedPackage" stuff to hope finding out where it has lost a
transitive dependency, also here absolutely no documentation what this
is supposed to do/work exactly beside some introduction that these xml
tags exits and reading the code... or probably add maven-compat
anywhere... or change provided to compile scope (even maven is jelling
at me that's bad and I will be punished soon)... not counting the many
times where I messed up the realms because I accidentally trying to use
XppDom objects in extensions and plugins and something between got
messed up.
With OSGi i get clear errors for missing requirements, I can clearly
share API (or declare I don't want to share it) and can reliable use it
without classlaoding problems.
If one wan't can even implement service filtering that would hide all
"illegal implemented API" ... and you can even make sure API is
(backward) compatible with implementation without waiting for a method
not found exception or alike.
Beside that I find it often more clear to distinguish between API (that
is only implemented by the framework) and SPI (that might be
implemented
by extenders). So probably it would be good to have maven-api and
maven-spi (instead of "maven-core") to make this clear.
Am 16.11.22 um 14:53 schrieb Romain Manni-Bucau:
Hi,
Guess classrealm is fine for maven, it does not bring much issues
(less
than OSGi or JPMS to be concrete), the real issue is the stability of
the
exposed API.
Thanks the hard work Guillaume did on that for maven 4 I guess it is
mainly
a matter of deciding what we do for maven 3.
Due to the resources and work needed I assume we can just play the
status-quo for maven 3.
Remaining question is for maven 4 do we drop the compatibilty. I
don't
like
much the idea but a compat layer can solve that smoothly for maven
= 4
and
limit the work needed, no?
Romain Manni-Bucau
@rmannibucau <https://twitter.com/rmannibucau> | Blog
<https://rmannibucau.metawerx.net/> | Old Blog
<http://rmannibucau.wordpress.com> | Github <
https://github.com/rmannibucau> |
LinkedIn <https://www.linkedin.com/in/rmannibucau> | Book
<
https://www.packtpub.com/application-development/java-ee-8-high-performance
Le mer. 16 nov. 2022 à 13:00, Christoph Läubrich <
m...@laeubi-soft.de>
a
écrit :
If you really like to separate API and get out of the
ClassRealm-Hell
OSGi would be much more suitable:
https://issues.apache.org/jira/browse/MNG-7518
Am 16.11.22 um 12:30 schrieb Gary Gregory:
As much as I dislike JPMS, maybe Maven 4 should migrate to Java 9
or
11
and
adopt JPMS to better define its public APIs.
Gary
On Wed, Nov 16, 2022, 05:06 Tamás Cservenák <ta...@cservenak.net>
wrote:
Yes, to define rules is quite easy, but to make our users to obey
them
is
tricky :D
In general, I guess, we should. For this reason JapiCmp has been
used in
Resolver since 1.9.0 (as noted on refd page end).
But while this was "kinda simple" to achieve in Resolver, I am
really
unsure if it is doable in Maven (sans 4 API) :(
Ultimately, this was the whole reason for API:
- users "grabbed" whatever could get hold on and used
- maven progress was really hindered by this, as that meant
modifying
(even
internal) interfaces without breaking clients was impossible, so
we
went
with tricks, and more tricks and even more tricks that now pays
back.
The other day we had a question on ML about 4-alpha compatibility
breakage,
and from mail it was clear that the package of the referred class
was
having "internal" in it. I mean, developers should really take
care
of
what
they import.
This is another huge plus for Takari lifecycle, it FORBIDS
compilation
against "encapsulated" internal classes....
http://takari.io/book/40-lifecycle.html#enforcing-dependency-usage-during-compilation
T
On Wed, Nov 16, 2022 at 10:44 AM Konrad Windszus <k...@apache.org
wrote:
I guess this is the easy part, the tricky question remains: Do we
need
to
make sure that all Maven 3 API interfaces/classes stay 100%
backwards
compatible until we reach 4.100/5.0/whatever?
This wasn't handled consistently in master till now, e.g. the
classes
generated from
https://github.com/apache/maven/blob/master/maven-plugin-api/src/main/mdo/lifecycle.mdo
are now immutable, i.e. lack setter methods in Maven 4.
My change in
https://github.com/apache/maven/pull/827/files#diff-2324c8cead0ad922c829a8ca450764aa149d6efdfe7f841e64210f20efd148acR77
was not backwards compatible (removed a method on an interface
which
may
have been implemented by extensions...)
Konrad
On 2022/11/16 09:35:15 Tamás Cservenák wrote:
Unsure we want to deprecate all of Maven :)
But yes, in general, 3.x "Maven API" was "all that users can
grab"
sadly,
and is probably our major source of problems and reason we
started
Maven
4
API.
IMO, I'd consider them as "whole", and just say "starting with
Maven
4.100/5.0/whatever" the maven-core (any class out of it) is NOT
ACCESSIBLE
ANYMORE FROM PLUGINS.
And done.
Just as an example, here is what Maven Resolver has to say about
same
topic
(part of not-yet-released, vote is in process 1.9.1 version):
https://maven.apache.org/resolver-archives/resolver-LATEST/api-compatibility.html
HTH
T
On Wed, Nov 16, 2022 at 10:26 AM Konrad Windszus <
k...@apache.org>
wrote:
I see now there is already
https://github.com/apache/maven/blob/master/api/maven-api-meta/src/main/java/org/apache/maven/api/annotations/Provider.java
but to me the javadoc is not explicit enough. It should state:
Only
Maven
is allowed to implement/extend types with this annotation.
Konrad
On 2022/11/16 09:20:11 Konrad Windszus wrote:
Hi,
Unfortunately Maven 3 didn’t define a proper API. In effect
everything
somehow exposed through class loaders was considered API by
plugin/extension developers.
For Maven 4 a completely separate API was established in
package
“org.apache.maven.api”, but what about the old packages used
and
exported
in Maven 3?
For example in the context of
https://issues.apache.org/jira/browse/MNG-7588 <
https://issues.apache.org/jira/browse/MNG-7588> I want to
evolve
the
package “org.apache.maven.plugin.descriptor”.
We already figured out that this particular package (although
not
part
of the Maven 4 API) is used from both Maven Core as well as
Maven
Plugin
Tools, therefore this probably needs to stay backwards
compatible.
What about others like
https://github.com/apache/maven/blob/master/maven-core/src/main/java/org/apache/maven/plugin/MavenPluginManager.java
?
<
https://github.com/apache/maven/blob/master/maven-core/src/main/java/org/apache/maven/plugin/MavenPluginManager.java
?>
This interface should IMHO never been implemented outside
Maven
Core
but
in fact it was exposed to all plugins/extensions (via
https://github.com/apache/maven/blob/a6b1ebb1cd40ca4b288fdeb30c6d2460323aa25b/maven-core/src/main/resources/META-INF/maven/extension.xml#L40
<
https://github.com/apache/maven/blob/a6b1ebb1cd40ca4b288fdeb30c6d2460323aa25b/maven-core/src/main/resources/META-INF/maven/extension.xml#L40
).
There are three options coming to my mind:
1. Deprecate the interfaces we don’t consider API and create
new
ones
for Maven 4 which are not exported!
2. Modify the existing interfaces in a backwards compatible
way
(but
somehow add a marker that they should not be implemented
outside
core)
3. Modify the existing interfaces in a backwards compatible
way
(but
somehow add a marker that they should not be implemented
outside
core)
For all three options we somehow need to come up with a list
of
classes/interfaces currently being exported through the API
class
loader,
which should be considered private and agree on an
Annotation/Javadoc
for
that (something like
https://github.com/mulesoft/api-annotations/blob/40b258afeff6560241dee5001ed00f1deb392e47/src/main/java/org/mule/api/annotation/NoImplement.java#L29
<
https://github.com/mulesoft/api-annotations/blob/40b258afeff6560241dee5001ed00f1deb392e47/src/main/java/org/mule/api/annotation/NoImplement.java#L29
or https://wiki.eclipse.org/API_Javadoc_tags#The_New_Solution
<
https://wiki.eclipse.org/API_Javadoc_tags#The_New_Solution>
WDYT?
Konrad
---------------------------------------------------------------------
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
---------------------------------------------------------------------
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
--
------------------------
Guillaume Nodet
--
------------------------
Guillaume Nodet
