Tom Hobbs wrote:
In my mind, it's an attractive proposition to drop support for the older
JVMs and go straight into 1.5 - however I'm only one user and I wouldn't
have a problem if River did that.
I think this question should be asked on the River Users list as well. If
there are active users out there who monitor River updates and the lists
then getting their feedback might make our decision easier.
That's a good point Tom, this project does need more focus on users.
In this case it would be about offering the user less and asking if it's
ok.
The only thing that Java 1.5 offers in Proxy code and Service API are
some additional classes, not much good for API, but useful for some
smart proxy's needing concurrency perhaps. But this is already
possible. Once we take the users mind away from supporting Java 1.4 in
Service API, however, we have to explain the difficulties of using
Generics in Service API. For example you can use Specific generics
List<Apple>, but not bounded wildcards, like List<? extends Fruit>, or
List<T>
Personally I can't see any reason or benefit as to why client or
platform Service API should be compiled with anything later than jsr14.
Separately compiled code needs to have the same type casts at run time.
Generics suffer from erasure and are replaced by type casts.
It would of course nix any remaining hopes I had for a Java CDC release,
after my recent failure, I have been considering creating an RMI
optional package to support it (rather than removing the need for RMI,
distribute it instead) Oracle and IBM have long since stopped
distributing the RMI optional package. Investigations are that it may be
possible to load the optional package upon instantiation, similar to the
way the jsk-policy.jar is now loaded. CDC will become a larger Java
client platform than the PC due to BlueRay and Digital TV, these come
with network connections. The ServiceUI would be very suited to
interactive Television. Server Platforms need clients too.
Long term I think we should still be making the build modular,
regardless of any decision taken with java 1.4.
By taking the build modular, it makes life easier for a greater
developer pool, just like our new Merge model.
Peter.
Also, it'd be good to start getting into the habit of consulting users when
significant decisions like this are being discussed.
Release early and often. This decision can be delayed to a later release,
rather than delaying the next release to contain the results of this
decision.
On Tue, Sep 14, 2010 at 8:53 AM, Peter Firmstone <[email protected]> wrote:
Yes it's a little daunting isn't it. It would make development,
maintenance and compatibility easier once we bite the bullet.
Any volunteers?
I had planned to release before Christmas, perhaps it should be a
maintenance release of the last branch, until we get a modular build and
Java 1.4 to 5 migration sorted.
Or do we let the release slip until March 2011 (1 year release) and aim for
a modular build, supporting Java 5, or do we do both?
Peter.
Patricia Shanahan wrote:
What time frame are you thinking of for all this? I was a bit naive
thinking we could easily switch to 5.
In order to make TaskManager progress, I need to pick a runAfter parameter
type. While things are in flux, I think I'll go with Collection<Task>, so I
can work with the jsr14 hack.
If we go to 5 for the thread package, I can do a refactoring pass to
change to Iterable<Task>. There are a lot fewer runAfter declarations
because of the switch to an abstract class with a default implementation.
Patricia
On 9/13/2010 6:26 PM, Peter Firmstone wrote:
Some time back, we voted to include Java 5 language features.
We didn't vote to drop support for Java 1.4 at the time and I was
considering different options for ongoing maintenance support of Java
1.4.
The last release 2.1.2, still supports Java 1.4
What I'd like to consider is that the previous branch be used to
maintain support for Java 1.4 and our next release only support Java 1.5
onward. However an additional requirement of a distributed platform
release, is that it plays nicely with existing Java 1.4 installations,
which means smart proxy's still need to be compiled in Java 1.4 bytecode
(and perhaps Service API). Service API will be already installed in the
clients local classpath for Java 1.4 installations, but the smart proxy
code depends on Service API and won't compile if they're Java 1.5.
So while Serialization compatibility is maintained across releases, we
just need to have compatible bytecode.
Perhaps with maven provisioning there could be an opportunity for
clients to select platform compatible bytecode at some time in the
future.
I guess we need to figure out how to compile the majority of classes
using "1.5", while compiling proxy classes using "1.4".
It is possible I think, to specify different class paths in the client
jvm's of the test suite, so potentially we could test the current
distribution with clients using the previous release and Java 1.4, while
the server uses Java 5 or 6.
This might require a separate release of the Test Suite, this would be
useful to test binary compatibility also, very important in a
distributed environment.
This allows a migration path away from Java 1.4.
"jsr14" is a build hack.
To achieve this, I think we'll need to change the way we currently
package River, to be more modular along the lines Dan Rollo has
suggested.
This might also be an opportunity to break out the Service
Implementations into separate components or subprojects perhaps, each
depending on the platform and test kit and each having it's own tests.
Allowing one to develop on a small component of the platform and run
only relevant tests, for faster feedback. All components might be
assembled and run on Hudson.
Possible Components:
Service API. (Java 1.4)
Jini Platform (depends on Service API) (Java 5)
Jini libraries. (Java 5)
Jini Security Policy.
Test Kit
Jini Compatibility Tests.
Reggie
Mahalo
Outrigger
Phoenix
Fiddler
Norm
Mercury
A client developer then only needs the Platform and Service API jar's.
Then when someone creates their own djinn, they can add the services as
they require them, reducing complexity for new developers.
Regards,
Peter.
Patricia Shanahan wrote:
Found it. common.xml contains a macro definition, javac-cmd, which I
believe does the actual compilation. It defaults the -target parameter
to "jsr14". Changing it to "1.5" allows my Iterable construct to
compile, but should continue generation of java 1.5 compatible code.
"jsr14" was an interim mode for experimentation with the 1.5 features
on a 1.4 JVM. It does support for-each for a Collection, but not
Iterable, which was added in 1.5, which explains my problem.
Any class files that work on 1.4 should also work on 1.5, explaining
Peter's observation.
Should River continue 1.4 support? If yes, we need to test on 1.4, as
well as 5, 6, and potentially 7. Note that there may be a performance
price, such as less efficient autoboxing code. If no, we should change
the target default to "1.5".
Patricia
On 9/13/2010 3:38 PM, Patricia Shanahan wrote:
....
Here's a theory. There is some parameter (or parameters) we don't know
about being passed to javac as part of the ant all.build that affects
the compiler's personality. It is generating 1.5 compatible code, and
rejecting a construct that I think should be accepted.
...
On 9/13/2010 1:16 PM, Peter Firmstone wrote:
[java] -----------------------------------------
[java]
[java] # of tests started = 581
[java] # of tests completed = 581
[java] # of tests skipped = 21
[java] # of tests passed = 581
[java] # of tests failed = 0
[java]
[java] -----------------------------------------
[java]
[java] Date finished:
[java] Tue Sep 14 01:08:05 EST 2010
[java] Time elapsed:
[java] 19953 seconds
[java]
BUILD SUCCESSFUL
Total time: 332 minutes 40 seconds
These test results are from qa.run executing on Java 5, so I was
expecting some test failures, knowing that the test suite doesn't
compile with JDK 1.5.
