I waited for almost 2 days. Your mail came just shortly after my commit.
So I thought the dicussion was settled.
You are right though that I could have waited a little longer. I only
committed the first part though and I think that I incorporated all
feedback into the change.
So let me summarize the feedback and how I included it:
- There was only positive feedback about having only the
startup.properties and not the properties in the config.properties to
control the startup bundles. So I think we agreed on this.
- David gave the feedback that he would like the startup.properties to
be a flat list of bundles. While I don´t agree that this is a good idea
this is absolutely possible. I used mvn urls in
startup.properties and for the framework but you can use relative paths.
So if you have the startup bundles directly below system you can just
refer to them by file name in the startup.properties.
I also used an interface ArtifactResolver that allows to easily replace
this behaviour.
- Guillaume and you gave the feedback to not introduce new dependencies
and I did not
So functionally I mainly simplified and restructured the Main code. It
went down from about 950 lines of code to about 400. I think the code is
much more readable and maintainable now.
There is one nice new feature that is possible now. You can specify the
path to your local maven repo in bundle.locations. In this case karaf
can startup completely without a system directory and directly uses the
local repo. I think this is really nice for developing. As you do not
have to do the assembly step when changing the source of modules in
startup.properties.
So again sorry for committing too early but I worked really hard on this
the weekend. In the end I had a really nice solution that at this point
I think fitted into all the feedback.
Christian
Am 07.05.2012 00:46, schrieb Achim Nierbeck:
To my concern the way main and the loading of karaf worked was "Good
Enough for Now".
I didn't see any issues to change the way it was working. So yes if
something is good and properly
working, don't Touch it. You might break it.
Regarding committing your changes, I do find it disturbing, that you
start a discussion without
waiting for a result.
regards, Achim
Am 06.05.2012 22:28, schrieb Christian Schneider:
I did not have the intention to make this more complicated. I just
removed the other options.
So what exactly do you -1?
I already committed the first step of the implementation and of
course did not introduce any new dependencies.
For the next step I plan to simply read the feature file instead of
the properties file. I don´t think that I need the feature service
for that.
Of course that means that the framework feature would only allow the
list of bundles and the startlevel for each bundle. So basically the
same
that is supported in the startup.properties. Is that ok?
Christian
Am 06.05.2012 19:12, schrieb Achim Nierbeck:
Even though you and Christian are certainly right that maven and
OSGi work quite well if the versions are kept right, but this
isn't the focus here. So coming back to the initial question I agree
with Guillaume, to better keep the main class
lean and simple therefore I give a -1 on this.
I don't want to see any dependencies to a features service what so
ever in main.
Thanx, Achim
Am 05.05.2012 11:33, schrieb Jean-Baptiste Onofré:
Agree with Christian.
Leveraging Pax URL in Karaf is a key feature (even if sometime we
fake the Pax URL usage, like in startup.properties URLs).
Regards
JB
On 05/05/2012 08:52 AM, Christian Schneider wrote:
Well in this case you should use felix it uses a flat list of
bundles :-)
I think the maven centric aproach is the biggest benefit in karaf. Of
course obr can help to make this even better but out there you almost
find no obr repos.
The big benefit with maven is that you have almost any lib available.
You only need to know the artifact coordinates. For example it is
great
that you can install cxf or camle by just
issuing two commands. How should that work without features that load
artifacts from maven? As soon as all bundles are available in obr
repos
we can switch to this aproach but
I think that is not the near future.
I think the aproach of installing features and bundles from a company
maven repo should be our recommended way of installing
applications. I
recommend to companies to split
their development and deployment process at the maven repo.
Developers
build the sources and deploy the binaries to the company maven repo.
Admins install from there. I think
that is the cleanest technical aproach to devops we currently have.
Of course this should include the use of the obr. As obr and maven
often
are incorporated in the same repository (like nexus or archiva) this
should be achievable.
Kar files are a dead end for me. They have their purpose when
companies
do not have a repository but they are completely anti modular. If you
deploy two applications using kar files you have a lot of duplication
and most of the advantages of osgi are gone.
About the flar system repo. Why should that be a good idea? The good
thing about the system repo as a maven repo is that it mimics the
central repo. So users can be sure that our system repo is just a
cache.
All the artifacts in there are the same as in central. So the user
knows
that each of these jars is the "official" version. This is very
helpful
for example for doing licensing audits.
Btw. I think maven and osgi are very compatible on the lowest level.
Maven can supply single artifacts very well. It is only the
dependency
resolution that is not compatible but obr can help out with that.
Christian
Am 05.05.2012 04:04, schrieb David Jencks:
I think we should make karaf much less maven centric including:
-- system repo is flat, not maven structured, with file names
enforced
as bundle-symbolic-name_bundle-version.jar. startup.properties can
then just have jar-name=start-level.
-- kar files use similar flat internal repo
-- non-kar features deprecated
-- heavily encourage use of obr.
maven and osgi are really not very compatible and trying to pretend
they are IMO leads to too many problems and suppresses the
usefulness
of osgi.
thanks
david jencks
On May 4, 2012, at 12:46 PM, Guillaume Nodet wrote:
Please, keep the main file lean and simple, no dependencies on url
handlers
or features or OBR or anything.
The less interactions we have with the framework, the less fixes
we'll to
do there and the more stable it will be.
The idea is to bootstrap the osgi framework, all the real
provisioning
should be done in the osgi framework itself using the feature
service or
obr or anything else that is required.
On Fri, May 4, 2012 at 8:35 PM, Jean-Baptiste
Onofré<[email protected]>wrote:
It makes sense, and I don't want to use the
OfflineFeatureService (not
require) but we will certainly have to decide to some
"restriction"
(for
instance, what do we do if a feature is define in a feature ;)).
Regards
JB
On 05/04/2012 08:18 PM, Christian Schneider wrote:
Hi JB,
yes we do not use the real maven resolution. I thought about
changing it
but it would have too many dependencies.
I did not mean to really use features. Rather to read the
feature file
instead of the startup.properties but still process and
resolve in the
same way as before. So this should not add
much complexity. We could use the OfflineFeatureService but I
dont
think
it is really necessary.
Christian
Am 04.05.2012 19:24, schrieb Jean-Baptiste Onofré:
Hi,
As reminder, in startup properties we don't really use mvn
URL. I
mean
we construct a file URL from the mvn one, we don't really use
Pax
URL.
Anyway, it sounds good to me. I don't think users use
anything else
than the startup.properties.
Regarding a feature instead of startup.properties, it means
that we
have to load at least feature core. I'm not sure that it's a
good
idea
because feature is already OSGi oriented, whereas in the main
area we
start the framework (so we are not in the "OSGi area"). It's
possible
but it means that even if we provide a features XML, it's not
really
the feature service that will be use but a FeatureStartup
process
(like OfflineFeatureService that we use in the Karaf maven
plugin).
So it means that we will have a dual bootstrap process which use
feature:
- the "startup" feature (which doesn't really use the feature
service)
- the "boot" feature (which uses the feature service)
As the startup.properties is generated from a feature
currently, it
makes sense to directly use the feature.
All depends the way that it will be implemented, but
basically +1
Regards
JB
On 05/04/2012 07:03 PM, Christian Schneider wrote:
Hi all,
on startup we currently use the following procedure.
We read property karaf.auto.start from the file
config.properties.
This can be either a list of bundles separated by spaces or
"startup.properties" or "all".
If it is all we replace karaf.auto.start with the list of all
bundles in
the system dir. I think this option does not really make
much sense.
If it is startup.properties then we replace karaf.auto.start
with
the
list of bundles specified in the file startup.properties.
Additionally we either support mvn urls or paths which are
converted to
mvn urls.
This all is quite a lot of variability of which we use none.
I propose to replace this in two steps:
1. Remove the karaf.auto.start property and always load the
bundles from
startup.properties. Also only support mvn urls.
This makes the code in main cleaner and makes it easier for our
users to
understand how to change the startup bundles.
2. Remove the startup.properties and instead use a feature
name to
determine the list of bundles to load
The second step makes this even simpler and additionally we can
remove
the generation of the startup.properties in the karaf maven
plugin.
So what do you think?
Christian
--
Jean-Baptiste Onofré
[email protected]
http://blog.nanthrax.net
Talend - http://www.talend.com
--
------------------------
Guillaume Nodet
------------------------
Blog: http://gnodet.blogspot.com/
------------------------
FuseSource, Integration everywhere
http://fusesource.com
--
Christian Schneider
http://www.liquid-reality.de
Open Source Architect
Talend Application Integration Division http://www.talend.com
