[SPARK-15857] Add Caller Context in Spark

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Assignee: Unassigned
• Reporter: Weiqing Yang
• Created: 10/Jun/16 00:37
• Updated: 26/Mar/20 12:40
• Resolved: 13/Feb/17 21:28

Description

Hadoop has implemented a feature of log tracing – caller context (Jira: HDFS-9184 and YARN-4349). The motivation is to better diagnose and understand how specific applications impacting parts of the Hadoop system and potential problems they may be creating (e.g. overloading NN). As HDFS mentioned in HDFS-9184, for a given HDFS operation, it's very helpful to track which upper level job issues it. The upper level callers may be specific Oozie tasks, MR jobs, hive queries, Spark jobs.

Hadoop ecosystems like MapReduce, Tez (TEZ-2851), Hive (HIVE-12249, HIVE-12254) and Pig(PIG-4714) have implemented their caller contexts. Those systems invoke HDFS client API and Yarn client API to setup caller context, and also expose an API to pass in caller context into it.

Lots of Spark applications are running on Yarn/HDFS. Spark can also implement its caller context via invoking HDFS/Yarn API, and also expose an API to its upstream applications to set up their caller contexts. In the end, the spark caller context written into Yarn log / HDFS log can associate with task id, stage id, job id and app id. That is also very good for Spark users to identify tasks especially if Spark supports multi-tenant environment in the future.

Issue Links

• breaks: SPARK-17710 ReplSuite fails with ClassCircularityError in master Maven builds
• is blocked by: HDFS-9184 Logging HDFS operation's caller context into audit logs
• is blocked by: HADOOP-13527 Add Spark to CallerContext LimitedPrivate scope
• is related to: SPARK-17714 ClassCircularityError is thrown when using org.apache.spark.util.Utils.classForName
• is related to: FLINK-16809 Support setting CallerContext on YARN deployments
• links to: [Github] Pull Request #14312 (Sherry302)
• links to: design doc