Hi,
I'm catching up the discussions and I have to admit the thread has too many
levels of nesting and it becomes difficult to read. Here is my view:
1) The structure of the distribution:
I see this as a repository of the Tuscany modules, 3rd party jars and other
things for the Tuscany standalone runtime and other environments where Tuscany
is embedded. IMO, we should keep the structure simple and independent of the
grouping. I'm happy with the current approach in 2.x:
modules
+--- tuscany-assembly-2.0.0.jar
+--- tuscany-assembly-xml-2.0.0.jar
+--- stax-api-1.0-2
+--- META-INF/MANIFEST.MF
+--- stat-api-1.0-2.jar
I'm against adding extra structures for the libs folder to achieve the
grouping.
2) Grouping of modules into functional features
This should be a very light layer on top of 1). It should be just a set of
configurations to organize the modules by function. When another distribution
is downloaded, it should be possible to just unzip it into the existing
distribution without any conflicts.
For the 1st milestone of 2.x, we could try to get "api", "core" and potentially
"webservice" distributions out.
3) Configuration of the features for specific hosting environments
* Maven: A pom project that list the set of modules for the functional feature
* JSE/WebApp: generate a launcher jar whose META-INF/MANIFEST.MF that uses the
Class-Path header to define the classpath entries
* OSGi: Generate a config.ini which lists the set of bundles to be used by the
OSGi runtime
* Eclipse: Generate a target platform definition or feature.xml
Thanks,
Raymond
From: Simon Laws
Sent: Monday, January 19, 2009 7:25 AM
To: [email protected] ; [email protected]
Subject: Re: [2.0] Align samples with the distributions
On Sat, Jan 17, 2009 at 8:53 AM, ant elder <[email protected]> wrote:
On Fri, Jan 16, 2009 at 2:01 PM, Simon Laws <[email protected]> wrote:
On Fri, Jan 16, 2009 at 1:04 PM, ant elder <[email protected]> wrote:
On Fri, Jan 16, 2009 at 12:28 PM, Simon Laws <[email protected]>
wrote:
snip....
Yes that what the current compact distribution does, its in 1.x right
now as there's few extensions going in 2.x but there should be no problems with
doing it there too.
...ant
I looked at the compact distro in 1.x and the distro projects in 2.x
and this is what I found:
Key
Plain text = common between the two
[] = different between the two
* = my opinion
A set of maven modules which define dependencies for features
[in /modules or /distribution]
* I prefer /distribution as the feature modules are not providing
any new function
There's pros and cons, where they are right now enables a facility i've
not mentioned yet but the where they go is not big deal at the end of the day.
Samples depend on only the feature modules that they require
An "all" distribution that ships all features
[separate distributions that ship individual features]
* I don't mind having them if people want to support them.
The problem with doing that is it adds some significant complexity, from
things like the download page, to things like the documentation for samples and
how to use etc.
As I say below I don't think we need to decide for M1. I would assert that
we need to be able to have samples depend on separate feature distributions,
modules call then what you will. Doesn't mean we need to ship them but does
mean they need to exist. As in my first point I think this is a common feature
of both approaches as they are currently implemented.
I don't think that would work very well, the "feature" approach is not a
feature per extension type so its only really useful for making downloadable
distributions. That means for example a sample using xquery you have to know
xquery is in the "process" feature distribution and download that, but you
wouldn't put a maven dependency on the process feature in the xquery sample as
that would drag in all the other unnecessary dependencies like bpel and ode,
same thing for say a jsonrpc sample which would have a dependency on the
jsonrpc binding not the web20 feature as that would drag in all the scripting
stuff. The feature idea was that there would be a small number of downloadable
feature distributions as a way to manage the growing number of Tuscany
extensions, the problem is though that there are no obvious groupings for most
of the extensions so the features get quite arbitrary as more extensions are
include which makes them unintuitive which just further complicates an already
complicated project. Since they were proposed we've asked on the dev list, user
list, and in the user survey and pretty much everyone is happy with the one big
distribution, i think a 70Meg or so download is just not a big deal these days.
What people do seem to want is less jars and an easier way to know which jars
are for what, and the compact jars and structured lib folder do that. Even if
we didn't use the compact jars, the structure lib folder makes a single
distribution much more usable, its obvious what jars are for and users can
easily make up their own custom distributions if required by just deleting
unneeded lib folders.
So we seem to be getting to the heart of the matter. We tentatively agree;
1 - start with one "all" distro
2 - add some structure to the libs dir
3 - have some grouping of jars that group modules together for easy reference
from samples etc.
On point 3 the disagreement is the granularity of the grouping. Features vs
extensions/compact.
Features = small number of features that group more that one extension
Extension/compact = larger number grouping each extension implementation
I'm not against having a finer granularity of feature
* I suggest though in this first instance (M1) we just ship the "all"
distro and we can see how we get on with that. This is still build from the
separate feature modules.
Yes i agree with just shipping an "all" distro for now. I'd like to leave
the "how its built" till this discussion has progressed a little further though.
[separate tuscany jars vs feature jars]
* I don't particularly like the feature jars as it's another step to
go wrong. In particular it uses the shade(?) plugin and you have to configure
transformers.
I'm puzzeled that you consider that as "another step to go wrong" but
don't mind having multiple different distributions or building distributions
out of each other, both of which seem to add much more complexity,
restrictions, and points of failure to me :)
To me it looks like a step in the process that we don't necessarily need.
In the context of this point my measure of complexity is how many places I have
to specify things. I agree that there are also complexities in distributing
more that one distribution package but I don't see that that's particularly
related to this point.
I'm still not understanding the problem here, adding a shade transformer
config is a pretty easy task, we've been doing it for years in the past Tuscany
releases and i don't remember a single occasion where its caused any problem.
And its only an occasional task for a Tuscany "developer" which makes the
"users" life easier every time they use Tuscany - and users have complained
about Tuscany having so many jars on lots of occasions. Also, if thats the only
hurdle in doing this then all we need to do is write a relatively trivial
Tuscany specific extension of the shade plugin and we then wouldn't even need
that transformer configuration. We're getting quite a lot of Tuscany specific
plugins these days so one more is no big deal.
Personally I've had a lot of problems with it. 50% of the time it fails in my
environment (having sais that I haven't tried is since I moved to the Sun JDK).
Anyhow the utility of this plugin to us will be dictated by the shape we want
our distribution to take. If we have a single distribution then it seems we
need to ship separate module jars as the OSGi runtime is expecting to load
bundles. I guess we could write a manifest aggregator for the shade plugin but
why would we when we can ship the separate module jars and provide a manifest
jar to make manual classpath specification easier. Maybe I'm missing something
here.
[Launcher vs manually specified classpath comes into this]
* Manual classpath is easier with feature jars. Can we use a
different approach to support the manual classpath? Manifest jars for each
feature?
I think we need to make a bit more progress on the "Tuscany runtime
launching" thread before that can be answered.
Agreed
[structured libraries directory]
* I like this. It gives some sense of order to the distro lib
directory.
Me too. Comparing the distros we've had in the past and the different
varrieties being talked about now i think a structured libraries directory
makes the most significant improvement for making the distro easier to use in
the way users have been asking for.
Some other distro thoughts came to mind as I went through:
1/ Can we do something about 3rd party licenses? E.g. automate the
command line tools we have that check distro jars are represented in the
LICENSE file so that this happens automatically when distro is built. Also
it would be nice to have a tool that checks that module NOTICE/LICENSE
files match the requirements of the source that they hold. This last is manual
at the moment.
The way to do this would be to change to use the maven release plugin and
IANAL plugin which would vastly simplify our distribution build and release
process, i think it is where i think really we need to get to but it will put
some restriction on how we shape the distributions which is one of the reasons
i'm less keen on all these "feature" distros
I looked at this a while back and I agree it adds restrictions. I don't
remember the details so I'll have to read up. I seem to think it was early days
when I looked at it so it was pretty rough at that point.
I think its pretty good these days, I know Geronimo uses it ok.
2/ We need to follow through on the reason Raymond started this thread
originally and ensure that samples are tested in the ways that a user will run
then when the distribution is built.
3/ As per 2/ but for webapp samples. Can we re-use some of the work
that was done for webapps in 1.x itest?
I'm starting to think there should be seperate distributions for OSGi and
JSE/Webapps etc. Trying to munge all this into a single distribution that works
for all environments so far looks like its going to make the distribution a lot
more complex and less user friendly than it needs to be. You could consider
this part of the "modularity" exercise of 2.0 and not munging unrelated things
together. If we did that then we can test the samples included in the
distribution properly in the best way for the target environment.
what would the difference be between these distributions? I think the jars
and their organization would be the same. It comes down to how you launch/host
the runtime so probably some hosting/launching jars would come and go. Again we
need to do more on the launching thread for this point.
Probably need actual distributions to show all the differences but for
example the organization isn't the same - the OSGi distro requires the 3rd
party jars be each included in its own sub folder which is unnecessary for
other environments and would make things like "java -cp" a bit unwieldy.
However, if we did go for the one common distro with a structure lib folder
then this becomes less of an issue for Ant scripts and launchers.
I suggest we lay out the distribution file directory in some detail and try
review it. Is this what we're talking about
tuscany-sca-2.0
bin
tuscany.bat (is this a bat or a jar?)
tuscany-default.conf
tuscany-osgi.conf
...
lib
base
acitvation-1.1
activation-1.1.jar
META-INF
MANIFEST.MF
asm-all-3.1.jar
tuscany-assembly-2.0.jar
tuscany-core-2.0.jar
tuscany-contribution-2.0.jar
...
tuscany-base-manifest-2.0.jar
spring
commons-logging-1.1.1.jar
spring-2.5.5.jar
tuscany-implementation-spring-2.0.jar
tuscany-spring-manifest-2.0.jar
webservices
samples
calculator
tutorial
store
