Dear Hive developers,
I am Dan from the Oozie team and I would like to bring up the
hive-exec.jar vs. hive-exec-core.jar topic.
The reason for that is because as far as we understand the official
recommendation from the Hive team is to use the hive-exec.jar artifact.
However in Oozie that can end-up in a binary incompatibility.
The reason for that is:
* Let's say library A is included in the fat Jar.
* And library B which is using library A is also included in the fat Jar.
* Let's also say that library A's com.library.alib package is
relocated to org.apache.hive.com.library.alib,
meaning the com.library.alib.SomeClass becomes
org.apache.hive.com.library.alib.SomeClass
* So if B has a method like public void
someMethod(com.library.alib.SomeClass) then the signature of this
method will be changed to:
public void someMethod(org.apache.hive.com.library.alib.SomeClass)
* If Oozie is also using B directly meaning we'll have b.jar on our
classpath, but with the unchanged signature,
so when hive-exec tries to invoke someMethod then depending on
whether b.jar coming from us will be loaded first or hive-exec will,
we can end-up with a NoSuchMethodError is hive-exec tries to pass an
org.apache.hive.com.library.alib.SomeClass instance to the
someMethod which was loaded from the original b.jar.
Hence in Oozie a long time ago (OOZIE-2621
<https://issues.apache.org/jira/browse/OOZIE-2621>) the decision was
made to use the hive-exec-core Jar.
Now since the shading process actually removes those dependencies from
the hive-exec pom which are included in the fat Jar, we manually had to
add some dependencies to Oozie to compensate this.
However these dependencies are not used by Oozie directly and with the
growing features of hive-exec we had to repeat the same process
over-and-over which is a bit unmaintainable.
Today I'm writing to you to propose a long-term solution where basically
nothing would change in the generated hive artifacts, poms and the same
time we wouldn't have to manually declare dependencies in Oozie which
are not explicitly used by us.
The solution:
1. We would create a new module named hive-exec-dependencies which
would be a pom-packaging module without any Java source files.
2. All the dependencies declared in hive-exec would be moved to
hive-exec-dependencies.
3. We would make the hive-exec-dependencies module the parent of
hive-exec and with this hive-exec would still have access to the
same dependencies as before.
4. The maven shade plugin would still strip the dependencies from the
generated hive-exec pom which are included in the fat Jar.
5. And with a small maven plugin we'd change hive-exec's parent back
from hive-exec-dependencies to the root hive project in the
generated hive-exec pom file.
I have a change ready locally and it works as described above.
With this on the Oozie side we could add a dependency on
hive-exec-dependencies and hence all the required libraries which are
included in the fat Jar would be pulled into Oozie.
The next time a new dependency would be added to hive-exec-dependencies,
the Oozie build would pull it in automatically without us having to
explicitly declare it.
Please let me know what you think.
Best,
Dan
