[jdom-interest] Java 5 planning

[Rolf]

Hi Jason, All.

Jason Hunter wrote:
> First off, I want to thank Rolf for investing his time in exploring 
> the technical feasibility of a JDOM migration to a Java 5 baseline.  A 
> few other people have explored the area in the past as well, and it 
> feels like (especially with JDOM 1.1 now finished) we can look at this 
> in earnest.
You're welcome.
> What I'd like to do here is take a high level view and ask some 
> questions of our local Java 5 porting experts -- Rolf, Mattias, 
> Victor, Gregor, and any others with experience porting widely-adopted 
> projects to Java 5.
>
> 1. I think we can all agree the ideal scenario is to develop a drop-in 
> replacement, a new JAR which (so long as you're on Java 5+) would let 
> previously compiled code continue to work.  This would let us continue 
> development as usual with a new release (call it 1.2 or 1.5 or 2.0) 
> simply having a higher Java version requirement.  How far can we go 
> down the road of embracing Java 5 while still maintaining this?
>
> 2. A less ideal scenario would require any previously compiled code to 
> be recompiled against the new JAR, but wouldn't require any code 
> changes.  This will cause some confusion and effort but to an 
> acceptably small amount.  Rolf, you say your build requires a 
> recompile and in 99% of cases doesn't require code changes.  Where are 
> the gaps?  Can they be filled?
>
> 3. The fallback scenario is one where people will need to modify 
> existing code in order to use the new version.  This will cause some 
> pain, and should we follow this path I think we should look at what 
> steps we can take to mitigate it.  Automated upgrade scripts?  A new 
> package, like org.jdom5, so people can choose to use the old version 
> or the new?  Other possibilities?
>
> 4. Lastly, I'd like to hear from people on what specific advantages 
> they're looking for in an upgrade to a Java 5 baseline.
>
> My answers:
>
> I'll start with (4) because you need to know why before you know how.  
> I think people would most like to improve the return values of methods 
> to be more specific.  For example, getChild() should return Element, 
> getChildren() should return List<Element>, text.getParent() should 
> return Element, doctype.getParent() should return Document, and so on. 
> Are there other "big ticket" items?
>
> On (1) I suspect that without a source recompile we can't adjust the 
> method return types which is the core value of the upgrade.
>
> On (2) from Rolf's README it looks like there are two areas likely to 
> require code changes based on his port: the Filter.filter() change and 
> the AttributeType enum.  Both seem like things to do if you're already 
> requiring code changes but aren't things so valuable in and of 
> themselves that we should require people do a code change.  Are there 
> more?
>
> -jh-
>
My answers, also starting with 4:

4a. For me the biggest advantage of a Java5 baseline is to reduce 
"compiler warnings". I rely on strict rules in the Eclipse development 
environment that do everything from ensuring that all method parameters 
are documented in the JavaDoc tags, ensure that methods that are 
overridden have the @Override annotation, etc. This leads to code which 
is well structured, and complete. My 'lint checking' in eclipse is more 
pedantic than the warnings generated from just the javac compiler. By 
porting JDom to Java5 I can *correctly* eliminate many warnings in my 
code, and this improves the signal-to-noise ratio in the code problems. 
I feel that it is important to be able to compile code without compiler 
warnings. Other people have different programming styles, but for me, a 
good Java5 port of JDom will make my life 'cleaner'.

4b. The next advantage is to be able to use the generic return types of 
methods to 'simplify' the code readability. Let's face it, "for (Element 
kid : root.getChildren()) {...}" is just so simple.

1. While I agree that a pure drop-in replacement jar would be ideal, I 
think too much would be compromised to get there. Looking at my 
prototype port, I imagine it could potentially be done, but then the 
AttributeType ENUM would have to be reverted. This is not actually a 
"big deal" I guess, using String constants is not bad, and I only used 
Enum because it is 'cool'. I think there may be other issues though... 
the change I made to 'Filter' class would require a recompile because I 
changed a method signature from accepting Object to accepting 
Content.... this would not work with a drop-in replacement. So, I think 
it *may* be possible, but unlikely. There may be more effort spent 
building a compromised product to accomplish a lofty goal which, for me, 
seems unnecessary.

2. This is the approach I aimed for... make the migration simple, and a 
recompile will do. There are some issues in the prototype that may/will 
cause issues:
- I changed a couple of method signatures have parameters that were more 
closely typed than JDom 1.1, like the Filter class above, where I take a 
Content parameter instead of an Object. In JDom1.1 it would not have 
made sense to send in non-Content data anyway, but it would have 
complied. With the prototype, it will not compile. As a result of this 
same change, it is now no longer possible to build a filter which 
returns Document... but, is there anyone out there who has a use case 
for that? JDom 1.1 had reference in the JavaDoc for the Filter class 
that you could build a filter that matches Document objects, but I could 
not think of a use case, and I 'needed' to have a Content input type, 
not Object, so I decided to drop support for it. This may cause compile 
problems for a small fraction of people.
- Some people may have specific references to Attribute Types that in 
JDom1.1 are String values... think along the lines of someone having the 
following in their code: String attype = Attribute.ID_TYPE. This will no 
longer compile because in the prototype, Attribute types are of the 
Enumerated AttributeType, and the code would have to become: 
AttributeType attype = Attribute.ID_TYPE. I don't believe AttributeTypes 
are used by many people, and the impact would be low.
- people who have chosen to have their own implementations of the Filter 
interface will be pretty messed up with the port. Then again, this is a 
'major release', and a change of interface is OK, right? Similarly, 
people who have extended any of the classes in JDom will have issues 
with compile errors...
One issue already encountered is that some people have already modified 
their 'legacy' code to be 'generified'. Mattias, for example, has done 
the following to his applications:

List<Element> something = 
(List<Element>)XPath.newInstance("/some/xpath").selectNodes();

the above will compile with a compiler warning on this line, but, all 
the subsequent references to the 'something' list will have the benefits 
of generics. Unfortunately, this means that the 'new' version of JDom 
will, unless it exactly returns List<Element>, will cause compile 
errors.... for example, it is possible that a final version of the 
selectNodes() method will return List<? extends Content>, and even this 
will cause compile *errors* in Mattias's code.

So, a bigger problem with the 'recompile and it works' situation is 
perhaps not the true legacy code, but the code which is ported to Java5 
already but will have different generics.... For what it's worth, in my 
legacy code ports I just happen to have used the wild-card generics to 
access Jdom, like List<?> kids = root.getChildren(), and, as a 
consequence, I have no problems.... perhaps other people will be OK too.

When I said 99% of people should get away with recompiles only, I was 
not aware of the situations like Mattias's. I need to revise that number 
much lower now...

3. Substantially changing the method signatures is something that I 
tried to avoid in the prototype, but, based on subsequent discussions, 
it appears that there may be places where this would be advantageous. 
XPath has generated some interesting challenges, and one solution would 
be requiring a second parameter to XPath constructors which would be the 
class of the return type... like: List<Element> something = 
XPath.newInstance("/some/path", Element.class).selectNodes(); I have 
actually built this alternate mechanism in an unpublished version of my 
prototype, and it does work quite well, and also conforms to the 
standard practices...


Additional Issues...

Based on discussions earlier, it appears that XPath may be a real 
problem with Generics. The 'right thing' for XPath to do depends on 
Jaxen, and, since that is not 'generified' we are stuck. The reason 
Jaxen is not generified is pretty obvious, because it is a whole lot 
more complicated to impose generics on a method call like 
Xpath.newInstance("//@*")....

Victor has made suggestions for the Filter class which are likely to be 
less intrusive than the changes I made in the prototype, but no-one has 
actually explored his suggestions yet.

Conclusion...

I don't believe that we can get a good Java5 API without at least 
compile-time compatibility. I think to get the best API we will probably 
need to resort to having substantially changed API's.

I think there are going to be compile-time errors for some people 
regardless of which route we decide to take....

Rolf