[HIVE-17574] Avoid multiple copies of HDFS-based jars when localizing job-jars

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.2.0, 2.4.0, 3.0.0
• Assignee: Chris Drome Assign to me
• Reporter: Mithun Radhakrishnan
• Created: 21/Sep/17 17:37
• Updated: 26/Sep/23 12:01
• Resolved: 05/Oct/17 21:59

Description

Raising this on behalf of Selina Zhang. (For my own reference: YHIVE-1035.)

This has to do with the classpaths of Hive actions run from Oozie, and affects scripts that adds jars/resources from HDFS locations.

As part of Oozie's "sharelib" deploys, foundation jars (such as Hive jars) tend to be stored in HDFS paths, as are any custom user-libraries used in workflows. An ADD JAR|FILE|ARCHIVE statement in a Hive script causes the following steps to occur:

1. Files are downloaded from HDFS to local temp dir.
2. UDFs are resolved/validated.
3. All jars/files, including those just downloaded from HDFS, are shipped right back to HDFS-based scratch-directories, for job submission.

For HDFS-based files, this is wasteful and time-consuming. #3 above should skip shipping HDFS-based resources, and add those directly to the Tez session.

We have a patch that's being used internally at Yahoo.

Issue Links

• is duplicated by: HIVE-17974 If the job resource jar already exists in the HDFS fileSystem, do not upload!
• relates to: HIVE-27737 Consider extending HIVE-17574 to aux jars