[jira] [Comment Edited] (SPARK-32385) Publish a "bill of materials" (BOM) descriptor for Spark with correct versions of various dependencies

["Vladimir Matveev (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-32385?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17186846#comment-17186846 ] 

Vladimir Matveev edited comment on SPARK-32385 at 8/28/20, 11:37 PM:
---------------------------------------------------------------------

[~srowen] Sorry for the delayed response!

> This requires us fixing every version of every transitive dependency. How does that get updated as the transitive dependency graph changes? this exchanges one problem for another I think. That is, we are definitely not trying to fix dependency versions except where necessary.

I don't think this is right — you don't have to fix more than just direct dependencies, like you already do. It's pretty much the same thing as defining the version numbers like [here|https://github.com/apache/spark/blob/a0bd273bb04d9a5684e291ec44617972dcd4accd/pom.xml#L121-L197] and then declaring specific dependencies with the versions below. It's just it is done slightly differently, by using Maven's {{<dependencyManagement>}} mechanism and POM inheritance (for Maven; for Gradle e.g. it would be this "platform" thing).

> Gradle isn't something that this project supports, but, wouldn't this be a much bigger general problem if its resolution rules are different from Maven? that is, surely gradle can emulate Maven if necessary.

I don't think Gradle can emulate Maven, and I personally don't think it should, because Maven's strategy for conflict resolution is quite unconventional, and is not used by most of the dependency management tools, not just in the Java world. Also, I naturally don't have statistics, so this is just my speculation, but it seems likely to me that most of the downstream projects which use Spark don't actually use Maven for dependency management, especially given its Scala heritage. Therefore, they can't take advantage of Maven's dependency resolution algorithm and the current Spark's POM configuration.

Also I'd like to point out again that this whole BOM mechanism is something which _Maven_ supports natively, it's not a Gradle extension or something. The BOM concept originated in Maven, and it is declared using Maven's {{<dependencyManagement>}} block, which is a part of POM syntax. Hopefully this would reduce some of the concerns about it.


was (Author: netvl):
Sorry for the delayed response!

> This requires us fixing every version of every transitive dependency. How does that get updated as the transitive dependency graph changes? this exchanges one problem for another I think. That is, we are definitely not trying to fix dependency versions except where necessary.

I don't think this is right — you don't have to fix more than just direct dependencies, like you already do. It's pretty much the same thing as defining the version numbers like [here|https://github.com/apache/spark/blob/a0bd273bb04d9a5684e291ec44617972dcd4accd/pom.xml#L121-L197] and then declaring specific dependencies with the versions below. It's just it is done slightly differently, by using Maven's `<dependencyManagement>` mechanism and POM inheritance (for Maven; for Gradle e.g. it would be this "platform" thing).

> Gradle isn't something that this project supports, but, wouldn't this be a much bigger general problem if its resolution rules are different from Maven? that is, surely gradle can emulate Maven if necessary.

I don't think Gradle can emulate Maven, and I personally don't think it should, because Maven's strategy for conflict resolution is quite unconventional, and is not used by most of the dependency management tools, not just in the Java world. Also, I naturally don't have statistics, so this is just my speculation, but it seems likely to me that most of the downstream projects which use Spark don't actually use Maven for dependency management, especially given its Scala heritage. Therefore, they can't take advantage of Maven's dependency resolution algorithm and the current Spark's POM configuration.

Also I'd like to point out again that this whole BOM mechanism is something which _Maven_ supports natively, it's not a Gradle extension or something. The BOM concept originated in Maven, and it is declared using Maven's {{<dependencyManagement>}} block, which is a part of POM syntax. Hopefully this would reduce some of the concerns about it.

> Publish a "bill of materials" (BOM) descriptor for Spark with correct versions of various dependencies
> ------------------------------------------------------------------------------------------------------
>
>                 Key: SPARK-32385
>                 URL: https://issues.apache.org/jira/browse/SPARK-32385
>             Project: Spark
>          Issue Type: Improvement
>          Components: Build
>    Affects Versions: 3.1.0
>            Reporter: Vladimir Matveev
>            Priority: Major
>
> Spark has a lot of dependencies, many of them very common (e.g. Guava, Jackson). Also, versions of these dependencies are not updated as frequently as they are released upstream, which is totally understandable and natural, but which also means that often Spark has a dependency on a lower version of a library, which is incompatible with a higher, more recent version of the same library. This incompatibility can manifest in different ways, e.g as classpath errors or runtime check errors (like with Jackson), in certain cases.
>  
> Spark does attempt to "fix" versions of its dependencies by declaring them explicitly in its {{pom.xml}} file. However, this approach, being somewhat workable if the Spark-using project itself uses Maven, breaks down if another build system is used, like Gradle. The reason is that Maven uses an unconventional "nearest first" version conflict resolution strategy, while many other tools like Gradle use the "highest first" strategy which resolves the highest possible version number inside the entire graph of dependencies. This means that other dependencies of the project can pull a higher version of some dependency, which is incompatible with Spark.
>  
> One example would be an explicit or a transitive dependency on a higher version of Jackson in the project. Spark itself depends on several modules of Jackson; if only one of them gets a higher version, and others remain on the lower version, this will result in runtime exceptions due to an internal version check in Jackson.
>  
> A widely used solution for this kind of version issues is publishing of a "bill of materials" descriptor (see here: [https://maven.apache.org/guides/introduction/introduction-to-dependency-mechanism.html] and here: [https://docs.gradle.org/current/userguide/platforms.html#sub:bom_import]). This descriptor would contain all versions of all dependencies of Spark; then downstream projects will be able to use their build system's support for BOMs to enforce version constraints required for Spark to function correctly.
>  
> One example of successful implementation of the BOM-based approach is Spring: [https://www.baeldung.com/spring-maven-bom#spring-bom]. For different Spring projects, e.g. Spring Boot, there are BOM descriptors published which can be used in downstream projects to fix the versions of Spring components and their dependencies, significantly reducing confusion around proper version numbers.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org