On 11/30/06, Pete Robbins <[EMAIL PROTECTED]> wrote:
On 30/11/06, Pete Robbins <[EMAIL PROTECTED]> wrote:
>
>
>
> On 30/11/06, Jean-Sebastien Delfino <[EMAIL PROTECTED]> wrote:
> >
> > Pete Robbins wrote:
> > > Our current method of packaging and loading an extension is fairly
> > > simple:
> > > we load all schema and libraries in the extensions path. This has a
> > > number
> > > of problems.
> > >
> > > 1. An extension may consist of more than one library e.g.
> > > libmy_extension.so
> > > and libmy_extension_utils.so. Our current loading scheme will
attempt
> > to
> > > load both of these and may fail on the one that doesn't provide the
> > > extension initialize method. On MacOS the output of our build
produces
> > a
> > > libx.dylib and a series of symlinks to this called libx.0.dylib ,
> > > libx.0.0.dylib etc.. ur runtime loads ALL of these which doesn't
cause
> > > problems as they are all the sam library and just repeatedly
register
> > to
> > > handle the same requests. Very inefficient though!
> > >
> > > 2. Control of whether or not to load an extension library is
currently
> > by
> > > renaming the library so the runtime doesn't find it. An example is
> > > that we
> > > ship our python extension as libtuscany_sca_python.so.diabled. This
is
> >
> > > horrible and error prone.
> > >
> > > We could improve this by having a system configuration file that
lists
> > > the
> > > required extensions but the I like the self contained package
approach
> > > that
> > > we have now. I'd like to implement an improved scheme for packaging
an
> > > extension by introducing a per extension configuration file.
Something
> > > like:
> > >
> > >
> > > tuscany_sca_ws_binding.extension
> > >
> > >
> > > <scacpp:extension name="ws binding" enabled="true">
> > > <library name="tuscany_sca_ws_reference"/>
> > > <library name="tuscany_sca_ws_service"/>
> > > </scacppp:extension>
> > >
> > > So the package would look like:
> > >
> > > extensions/
> > > ws/
> > > tusany_sca_ws_binding.extension
> > > lib/
> > > xsd/
> > > other_folder/
> > > ...
> > >
> > > The .extension configuration file is saying to load the library
which
> > is
> > > located somewhere in the package... the runtime will find it... no
> > > need to
> > > specify a path.
> > >
> > > Taking this further the configuration file could list the schema to
be
> > > loaded. Currently the runtime will just load any it finds but these
> > > may not
> > > be needed by the runtime e.g. the schema may be for some extension
> > > implementation specific purpose.
> > >
> > > I think it would also be good for the extension initialization()
> > > method to
> > > take as a parameter the root of the extension e.g.
> > > extension("/tuscany/extensions/ws"). This would allow the extension
> > > package
> > > to contain any configuration information that it needs.
> > >
> > > I'd like to start by at least introducing the .extension file for
each
> > > extension and loading only the specified library(ies) if the
extension
> > is
> > > enabled.
> > >
> > > Any thoughts?
> > >
> > >
> > >
> >
> > Two thoughts:
> > - convention over configuration
> > - the runtime should be consumable without having to go tweak XML
> > configuration files
> >
> > If I remember correctly, renaming the dlls to .disabled was a last
> > minute change to work around DLL loading errors with our M2 release on
> > Windows.
> >
> > I agree that we should do better than renaming to .disabled, but I'd
> > like to understand better the actual issues that we faced before
> > inventing yet another XML based runtime configuration language :)
> >
> > I'm aware of the following issues:
> > 1. We need the runtime to load extension libs only, not other libs
under
> >
> > the extension directory which are not actual extensions
> > 2. Same for XSDs, we need to load XSDs that contribute to SCDL, and
> > leave other XSDs under the extension directory alone
> > 3. Extensions that cannot be loaded because some of their dependencies
> > are not there should no break the runtime
> > 4. The system admin / installer should be able to disable extensions
> > that won't load because their dependencies are not there (I'm not yet
> > convinced that this is still an issue if we manage to solve issue #3)
> >
> > Did we run into any other big issues?
>
>
> No, that's about it.
>
> 2. XSD loading can be done by convention (schemas for the runtime are
> always in a folder called 'xsd'
> 1. Could be solved in a similar way by only loading libraries in a
folder
> called ???
> 3. Can be solved by just ignoring the load errors/issuing a warning
rather
> than giving up
> 4. Can be solved by solution to 3.
>
>
>
I recall now why 3. was a big problem. Windows sometimes throws up an
error
dialog when the load fails so it was not just a case of the runtime
handling
the load failure so we had to "disable" the extension.
Cheers,
--
Pete
--
Pete
Pete, just summarizing so that I understand. It seems there are two parts
to this.
A/ Building and installing an extension in order that it can be consumed by
the runtime. Building and installing extensions is potentially a separate
activity from building and installing the runtime itself. Looking at the M2
download they all come together at the moment.
B/ Optionally consuming/enabling an extension in the runtime, once the
extension has been installed, in such a way that the runtime is able to find
and load it.
If I understand the previous posts (and the current code) A/ is achieved by
building the extension in the deployed folder layout
extension type/
bin
lib
xsd
and Pete you are proposing that this is adjusted so that there are
directories that hold just the stuff that the contains the extension rather
than the things it depends on.
extension type/
bin
extensionbin?
lib
xsd
extensionxsd?
You are also proposing that B/ is achieved by ensuring that this directory
structure be placed in a location that the the runtime can search for active
extensions (as it does at the moment in the deploy directory) and that the
runtime ignores any badly configured extensions if possible. Removing an
extension's directory from the deploy directory structure has the effect of
disabling it so I guess this is the fallback if the runtime can't continue.
Anything more complex and, as Jean-Sebastien suggests, you start getting
into the full blown package and dependency management problem that many
other systems try to solve in different ways, e.g. rpm or pear.
Simon
