bug#31721: “generate-jar-indices” phase fails on many packages

[Ricardo Wurmus <ricardo.wurmus@mdc-berlin.de>]

Danny Milosavljevic <dannym@scratchpost.org> writes:

>> The errors are either about duplicates as in java-picard-2.10.3:
>> 
>>     duplicate entry: htsjdk/samtools/AbstractBAMFileIndex$1.class
>
> That's... bad.  What if those files differed?  Which one would be used at
> runtime?

I wonder where they come from.  Is this from a dependency or are there
actually two such classes?

>> or about missing files as in dropseq-tools:
>> 
>>     java.io.FileNotFoundException: /gnu/store/q76y0ximcziplgfpbn26kbw4h3s14f33-dropseq-tools-1.13/share/java/lib/biojava-alignment.jar
>
> That's also bad.  How does it run at all if its dependencies are missing?

To satisfy the dropseq-tools build system the dependencies a symlinked
to the build directory.  I suspect that the error is something to do
with that.

> The only thing I can think of that would improve things long term:
>
> The phase can add relative paths to all the dependencies to META-INF/MANIFEST.MF
> before invoking "jar -i".
> Then "jar -i" will index those - and all java packages can use regular inputs
> instead of propagated inputs.  I've tested that locally already - and it
> works fine.

Is it correct that this would ensure that in case of duplicate class
names this would pick the class from the current package?

-- 
Ricardo