Hi Geir,
I think that what Tim means is that we have to guard against *other
modules* having dependencies on internal APIs that are likely to change.
For instance, suppose ModuleA has a class MyThing in the
org.apache.harmony.modulea.internal.foo package that has been developed
with the intention of only being used by types in ModuleA. Placing class
MyThing in an "internal" package sends a very strong signal out that
this class should not be used outside of this module. It could be that
the class is prone to a lot of changes ; or perhaps the ModuleA
developers have agreed that at some point in the near future they will
replace the class with a new type. Either way, it is a "domestic" issue
that is for the developers of ModuleA to discuss and sort out among
themselves.
If a developer of *ModuleB* imports
org.apache.harmony.modulea.internal.foo.MyThing into their code then
they now run the risk of breaking ModuleB every time the ModuleA team
make changes which is pretty disastrous. And such breakages occur more
often than one would imagine - hence the appearance of things like this
: http://www.alphaworks.ibm.com/tech/aus .
That situation could be avoided by the ModuleB developer(s) respecting
the "internal" part of the package that they are thinking of importing.
If tooling can be used to enforce the rule of not using *other modules'*
internal classes then all the better.
Sorry if this is stating the obvious.
Best regards,
George
Geir Magnusson Jr wrote:
Tim Ellison wrote:
The problem is that once you release a module with a particular internal
API into the wild you need to ensure that other modules that depend upon
it are not broken by your internal API changes.
But aren't they internal?
In general, unless you
control the make-up of the runtime you will not know whether you are
breaking other modules (or to put it another way, you can check that you
don't break other modules for a given build but other people may be
combining different versions of the modules). With runtime support you
can have multiple versions of a module simultaneously, without runtime
support you must retain binary compatibility. In practice this simply
means knowing which changes are allowed, and which require a new
internal API.
Regards,
Tim
On 2/13/06, Tim Ellison <[EMAIL PROTECTED]> wrote:
I've put a first cut of the proposed package naming convention on the
website:
http://incubator.apache.org/harmony/subcomponents/classlibrary/pkgnaming.html
Regards,
Tim
Leo Simons wrote:
On Fri, Feb 10, 2006 at 12:52:09PM +0300, Anton Avtamonov wrote:
As I wrote elsewhere, I propose that packages whose naming
convention is:
org.apache.harmony.<modulename>.<something>
represent internal APIs. All visible (public/protected) types
in those
packages can be used by class library developers from any
module, and
such developers can expect the API to be evolved in a compatible
way.
Module developers should not rely on the stability of anything
starting:
org.apache.harmony.<modulename>.internal
and are strongly discouraged from referencing visible types in such
packages since these are type internal module implementation
code (and
when we turn on OSGi runtime checks the imports from other
modules will
fail).
From other hand if we are talking about using these useful
classes only
inside Harmony then it's probably a good idea. But we need some
procedure
for moving a class to utilities package and we need to notify
developers
about class capabilities.
Moving it to a 'utilities' package should be as simple as making
it a
non-internal package name; and the notification is javadoc of those
non-internal types.
If people agree I'll write something for the website along these
lines,
if not we can continue to debate it on the list.
+1 (to the website thing, though debate obviously is fine too :-D).
I personally think that it is a good idea to seperate out
non-J2SE-specific
non-Harmony-specific stuff as much as possible. I also think it is
a good idea
to evaluate very carefully on a case-by-case basis whether it
makes sense to
have that kind of code live under the "harmony" banner --
experience has shown
that breaking out "true" utility stuff can help result in wider
usage of that
stuff which tends to result in more bugfixes, better code, etc.
For example (and I'm being really naive here) I can imagine that
java.util
could be built on top of jakarta-commons-collections, or that
java.util.logging
could be built on top of log4j (now that'd be cool).
But this is just me imagining stuff that might be possible in the
future. For
now it doesn't make sense to worry or think about this too much -
we'll see
what kind of ties between harmony and something like
jakarta-commons (or
whatever other part of the open source java space we're on about
here) is
possible as the issue comes up.
cheers!
Leo
--
Tim Ellison ([EMAIL PROTECTED])
IBM Java technology centre, UK.
