OSGi import/export headers
ascertrobw opened this issue · comments
Realise this is a very niche/uncommon area these days. Would be handy though to have OSGi import/export declarations in the Manifest so that the library can be used in an OSGi runtime framework with other libs that create these. Can supply examples if it is of interested to be added.
I'm not sure we have someone amongst our maintainers that is an OSGi expert. I'm reluctant to these additions as we have not means to properly test such metadata nor do we have the expertise to maintains OSGi declarations. Such arrangements typically ask for constant trouble and maintenance.
Yeah - I get that, it is a bit of a fringe area these days. In fact, looking at what the BND tool produces, I don't think it's too hard in your case for just the SPI JAR. Here's how the Manifest comes out once we push it through the "auto BND" in our library re-build stage
Manifest-Version: 1.0
Bundle-Category: bundle
Bundle-ContactAddress: support@ascert.com
Bundle-Description: r2dbc-spi packaged for OSGi
Bundle-ManifestVersion: 2
Bundle-Name: r2dbc-spi
Bundle-SymbolicName: com.ascert.repackaged.r2dbc-spi
Bundle-Vendor: Ascert, LLC
Bundle-Version: 0.8.5.RELEASE
Export-Package: io.r2dbc.spi;version="0.8.5.RELEASE"
Import-Package: javax.annotation.meta;resolution:=optional,javax.annot
ation
Require-Capability: osgi.ee;filter:="(&(osgi.ee=JavaSE)(version=1.8))"
Most of the above is boilerplate our re-build creates - the active parts really are the Import-Package / Export-Package headers. We made the javax.annotation.meta optional which may not be strictly correct for your code, but since we only need the bundle to resolve (we're not using the underlying classes) it handles things our side.
Such arrangements typically ask for constant trouble and maintenance.
I've added these meta data to jOOQ early on using the Apache Felix Maven plugin, and surprisingly, I hardly ever had any need to maintain them, which can only mean two things:
- No one is using OSGi
- The meta data that most OSGi users need really is stable
OSGi is a rarer usage case for sure - I couldn't speak to the wider usage, but Apache Felix is certainly used and maintained which is how we use our OSGi framework. I believe the Eclipse IDE itself was OSGi framework based at one stage, although we mostly used Netbeans or VScode these days.
I guess the imports must have changed with jOOQ 3.15, since this issue only arises then. Presumably r2dbc-spi was either not an import in 3.14, or was an optional one.
Presumably r2dbc-spi was either not an import in 3.14, or was an optional one.
It was not
BTW - I'm happy to submit a pull request with the extra headers if of help. Not a MVN expert, as we're gradle users, but pretty sure it's a small change to the current pom.xml:
<configuration>
<archive>
<manifestEntries>
<Automatic-Module-Name>r2dbc.spi</Automatic-Module-Name>
</manifestEntries>
</archive>
The essentials as per the following example
Bundle-Category: bundle
Bundle-ManifestVersion: 2
Bundle-Name: r2dbc-spi
Bundle-SymbolicName: io.r2dbc-spi
Bundle-Version: 0.8.5.RELEASE
Export-Package: io.r2dbc.spi;version="0.8.5.RELEASE"
Import-Package: javax.annotation.meta;resolution:=optional,javax.annot
ation
Require-Capability: osgi.ee;filter:="(&(osgi.ee=JavaSE)(version=1.8))"
Optional, but not really required being:
Bundle-ContactAddress:
Bundle-Description:
Bundle-Vendor:
A PR sounds good, with having Bundle-Version and Export-Package / version=… parametrized it should be a rather small change. javax.annotation is only required during compilation and is mostly there for IDE support. Frameworks like Spring can additionally introspect API for nullability annotations, but I'm not sure whether there are OSGi-based tools for nullability introspection. Since all this can be an enhancement for existing pom config and we don't need a plugin here, it's less of an hassle than initially anticipated.
One thing to note, R2DBC SPI is a component that interacts with Java's ServiceLoader to load driver classes. Is this something to be considered in OSGi metadata?
Cool - I'll take a look. I'm not a Git expert, so forgive me if I fumble a bit on the local fork / PR creation process. Always have to remind myself on the GitHub process for that
One thing to note, R2DBC SPI is a component that interacts with Java's
ServiceLoaderto load driver classes. Is this something to be considered in OSGi metadata?
You have me there - I'll have to go digging into some Docs. Haven't looked into how OSGi and ServiceLoader co-operate and if additional metadata is needed.
Update: this document looks like the relevant spec, and seems to be involve extra manifest declarations on the provider side and consumer side. I'm guessing r2dbc will need the consumer side ones?
I'm guessing r2dbc will need the consumer side ones?
I think so. We call ServiceLoader.load(io.r2dbc.spi.ConnectionFactoryProvider, …).
Submitted pull request for the basic headers.
Might be handy if feasible to also merge into the 0.8.x release branch.
Thanks a lot, the PR looks pretty decent. Can you add the ServiceLoader entries as well? Once that's in place, we can merge the change for both, the main and 0.8.x branches.
Partly because I have no test case or environment in which I can prove they actually work - but also because I have a feeling they may well trigger knock on set of dependencies.
Exactly the same here, even worse, I never really got in touch with OSGi details.
I'd rather defer on the ServiceLoader entries for this commit if we can
We don't have many attempts to get things right. We should do it in a single step. I cannot oversee how much or little impact we will have with the metadata already so I'm inclined to wait with the addition until we can deliver the full set of metadata.
I'd rather defer on the ServiceLoader entries for this commit if we can
We don't have many attempts to get things right. We should do it in a single step. I cannot oversee how much or little impact we will have with the metadata already so I'm inclined to wait with the addition until we can deliver the full set of metadata.
Let me see if OSGi supports the concept of an "optional" ServiceLoader requirement on the consumer side. If not "right" becomes a bit debatable. With the present headers, the bundle will load and not prevent the OSGi framework from starting if no service providers are present. That would be right for any bundle that needs the SPI classes to resolve, but may not actually use the underlying service - which often times with OSGi SPI or API bundles, is as many or more cases than those who use the provided service.
OK - added an extra commit to the PR for the ServiceLoader headers.