I'm checking with a following method for judging whether user-facing api use Guava's object, though it's not elegant one:

1. Executing find . -name "*.java" | xargs grep "google.common" under libraries which seems user-facing api (hadoop-hdfs-project/hadoop-hdfs-client/, hadoop-common-project, , hadoop-yarn-project/hadoop-yarn/{hadoop-yarn-client, hadoop-yarn-client}, hadoop-mapreduce-project/hadoop-mapreduce-client, )
2. Open code and whether it has @InterfaceAudience.Private. If so, we don't need to check the code.
3. Check the usage of guava's objects.

I checked hadoop-hdfs-project/hadoop-hdfs-client/, and there are no exposure of Guava's object.

Since HADOOP-10101 was committed, is the plan to finish this investigation, then shade Guava everywhere?

A more sure way of finding API pollution would be writing a bytecode analysis tool. We'd filter for classes and methods annotated IA.Public and LimitedPrivate, then check those method signatures for Guava class references.

Another possible quick thing you could try is using JACC or JDiff to compare current vs. shaded Guava.

Thanks Andrew for the information. Annotation File Utilities, being included in Checker Framework, can extract Annotations included in classes. Hence, we can use it to check whether classes have IA.Public and LimitedPrivate. Let me try.

https://checkerframework.org/annotation-file-utilities/#extract-annotations

I'm upgrading this to a blocker, since it's a prereq for shading. I'm also going to file a blocker JIRA for shading guava everywhere.

Hi Tsuyoshi Ozawa, do you have an update for this JIRA?