This is an interesting question and it usually has consequences that are
far-reaching in user experience.
If a Flink app is supposed to be a "standalone app" that any Flink
installation should be able to run, then the child-first classloading makes
sense. This is how we build many of the Java application servers (e.g.
GlassFish, JBoss etc). Doing this makes the application "self-contained"
and perhaps portable. Of course, this increases the size of the Jar. The
one issue to watch out for is application using framework classes that are
newer than framework itself. For instance, should I expect my app with
Flink *1.6* DataSet/DataStream classes to run smoothly on a Flink 1.5
If a Flink app depends on a particular (version of the) Flink installation,
then, if using parent-first classloading, the app can make use of the
classes that the installation itself uses. This makes the app
(comparatively) less self-contained, but this limits the size of the app's
Jar. There are advantages of doing this, but it poses problems especially
Whether one or the other should be the behavior largely depends on how the
applications are built, tested, and deployed. Application's build comes
into picture because in tools like Maven a dependency can be declared to be
"provided" which means if you know that your app's dependency is also your
framework's (i.e. Flink) dependency and you, as an app developer, are okay
with that Maven wouldn't bundle it in your app's Jar.
So, my recommendation is that since this appears like a backward
incompatible change, Flink should provide an option to go back to
parent-first classloading for a given app, at least for 1.5. Child-first
classloading seems like the right thing to do given how (unnecessarily)
complicated the deployments have become and given how frequently apps use
library versions that are different from the framework.
ElasticSearch solution has merits too, but it is unclear if it helps *at
deployment time* merely to identify that there is a duplicate (without
knowing where it has come from). Ideally, when people build the so-called
shadow Jar (one Jar with all dependencies) the build script should warn of
the duplicates. Shadow Jars alleviate (but do not remove) the problems of
"Jar Hell". But it seems to me that till we move to a modular Java (that is
Java 9; I think this is way out in future), this is the preferred solution.
That said, I'd really like to see a classloading section in Flink docs
(somewhere in dev/best_practices.html). Is a JIRA in order?
On Fri, Mar 9, 2018 at 1:52 PM, Stephan Ewen <ewenstep...@gmail.com> wrote:
> @Ken very interesting thought.
> One for have three options:
> - forbid duplicate classes
> - parent first conflict resolution
> - child first conflict resolution
> Having number one as the default and let the error message suggest options
> two and three as options would definitely make users aware of the issue...
> On Fri, Mar 9, 2018, 21:09 Ken Krugler <kkrugler_li...@transpac.com>
>> I can’t believe I’m suggesting this, but perhaps the Elasticsearch
>> “Hammer of Thor” (aka “jar hell”) approach would be appropriate here.
>> Basically they prevent a program from running if there are duplicate
>> classes on the classpath.
>> This causes headaches when you really need a different version of library
>> X, and that’s already on the class path.
>> See https://github.com/elastic/elasticsearch/issues/14348 for an example
>> of the issues it can cause.
>> But it definitely catches a lot of oops-ish mistakes in building the
>> jars, and makes debugging easier (they print out “class X jar1: <path to
>> jar> jar2: <path to jar>”).
>> Caused by: java.lang.IllegalStateException: jar hell!
>> class: jdk.packager.services.UserJvmOptionsService
>> — Ken
>> On Mar 9, 2018, at 3:21 AM, Stephan Ewen <se...@apache.org> wrote:
>> Hi all!
>> Flink 1.4 introduces child-first classloading by default, for the
>> application libraries.
>> We added that, because it allows applications to use different versions
>> of many libraries, compared to what Flink uses in its core, or compared to
>> what other dependencies (like Hadoop) pull into the class path.
>> For example, applications can use different versions of akka, Avro,
>> Protobuf, etc. Compared to what Flink / Hadoop / etc. uses.
>> Now, while that is nice, child-first classloading runs into trouble when
>> the application jars are not properly built, meaning when the application
>> JAR contains libraries that it should not (because they are already in the
>> classpath / lib folder).
>> For example, when the class path has the Kafka Connector (connector is in
>> the lib directory) and the application jar also contains Kafka, the we get
>> nasty errors due to class duplication and impossible class casts (X cannot
>> be cast to X).
>> What I would like to understand is how this change worked out for the
>> users. Based on that, we can keep this or revert this change in the next
>> Please answer to this mail with:
>> a. This was a great change, keep it and polish it.
>> b. This caused in the end more problems than it solved, so please set
>> the default back to "parent-first" in 1.5 and leave "child-first" as an
>> optional flag.
>> Thanks a lot,
>> +1 530-210-6378 <(530)%20210-6378>