Github user steveloughran commented on the issue:
https://github.com/apache/spark/pull/12004
(apologies for not replying; rebuilding a deceased laptop)
My main concern is to have the ability to make spark releases which include
the object store client libraries and a set of transitive JARs consistent with
the version of Hadoop and spark used. It's that transitive problem, "consistent
aws-sdk" and "all jackson JARs in sync" which makes it hard and stops it being
straightforward for any downstream project to pull in the right files
themselves. I know this, because I have code that wants to do that, and,
because I'm testing across so many different variants of the hadoop-* modules,
I get to see these things first. Put differently: this patch compensates for
the fact that whenever I bump the dependency version of the aws- or azure- JARs
spark apps trying to work with these object stores break.
With the packaging set up, then anyone can build spark itself with the
right JARs, including the (large) transitive AWS/Azure dependencies. Pretty
much everybody publicly releasing derivatives of spark are doing this âan
explicit module delivers that same ability to the ASF code itself. And: allows
the spark project to publish to maven central, the spark-hadoop-cloud artifact
to let anyone building apps downstream via maven, sbt, ivy, ... to pick up the
right dependencies. Trying to do that downstream is a very delicate piece of
work.
I'm about to push up a version which has just cut out the tests for
transitive classloading; the means there's no source in the module, just the
packaging. It is left to downstream code to validate the artifact declarations
through whatever functional tests they have. It still generates a JAR file,
only this is now empty. I could change it to being a POM artifact only, though
that would commit the module to be a POM-only artifact forever.
Now, once the code is stripped down to its minimum, there is one more
deployment option: just adding the artifacts to an existing module (e.g
spark-core), again presumably with some profile to enable it. That's actually
simpler: no new spark artifacts, just tuned dependencies. spark-core already
ships with the jets3t dependency, because hadoop-common (still) declares its
dependency on it. Happy to do it that way if you want : all I care about is
having the packaging and transitive dependencies available and consistent.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
