[HUDI-3763] Fixing hadoop conf class loading for inline reading

[nsivabalan]

What is the purpose of the pull request

If we enable point look ups for Hfile in MDT(hoodie.metadata.enable.full.scan.log.files), it fails w/ class not found. Check the linked jira for details. We did test out inline FS before. but one of the commit overwrote it.
Commit where this was flipped: https://github.com/apache/hudi/pull/4333/files#r840216398
So, this patch fixes the hadoop conf for inline FS.

Brief change log

Fixed hadoop conf for inline FS.

Verify this pull request

• Verified quick start guide works w/ local FS and S3.

Committer checklist

• Has a corresponding JIRA in PR title & commit

• Commit message is descriptive of the change

• CI is green

• Necessary doc changes done or have another open PR

• For large changes, please consider breaking it into sub-tasks under an umbrella JIRA.

[vinothchandar]

but this wont work. new Configuration() will lost everything else set on the driver

[vinothchandar]

We have to ensure we use the hadoopConf object passed down from the engine

[nsivabalan]

sure. will get to the bottom of it to see why wiring it fails.

[nsivabalan]

found the fix. have updated it.

[codope]

This is very weird. I can clearly see the class is packaged in jar:

jar tf packaging/hudi-spark-bundle/target/hudi-spark3.2-bundle_2.12-0.11.0-SNAPSHOT.jar  | grep -i inlinefile
org/apache/hudi/common/fs/inline/InLineFileSystem.class

and yet the write fails with ClassNotFoundException: Class org.apache.hudi.common.fs.inline.InLineFileSystem not found. So, it's not that the class is missing. Something else is overriding the classpath due to which classloader is unable to pick it.

Just to verify, I manually placed the hudi-spark-bundle under <SPARK_HOME>/jars and then tried the same script. It ran fine and upserts went through.

[codope]

fyi, tested with 0.10.1 and it works fine.

[codope]

Some more info:
i used following to print classpaths just after starting spark-shell, referred https://stackoverflow.com/questions/30512598/spark-is-there-a-way-to-print-out-classpath-of-both-spark-shell-and-spark

import java.lang.ClassLoader
val cl = ClassLoader.getSystemClassLoader
cl.asInstanceOf[java.net.URLClassLoader].getURLs.foreach(println)

This does not show the path to my hudi jar. Maybe, something overrides in the bootstrap step itself. But then, 0.10.1 works so that means it's got to something to do with change in Hudi.

[nsivabalan]

thanks @codope for the assistance. fixed it now.

[codope]

Siva and I verified the patch. This is the way to retain hadoop conf as well as add any other conf where we want to set hudi class. We do like this in SparkHelpers too.

[hudi-bot]

CI report:

• f6a68c0 Azure: SUCCESS

Bot commands

@hudi-bot supports the following commands:

• @hudi-bot run azure re-run the last Azure build