[SPARK-22336] When spark-submit cluster mode is run from a Mesos task, the job fails

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.0.2
• Component/s: Mesos
• Labels: bulk-closed

Environment

Mesos

• Assignee: Unassigned
• Reporter: Susan X. Huynh
• Created: 23/Oct/17 15:48
• Updated: 21/May/19 04:14
• Resolved: 21/May/19 04:14

Description

In Mesos cluster mode, spark-submit passes all local MESOS_xxx env vars to the Dispatcher (via RestSubmissionClient). When these env vars are later set in the Driver task, they overwrite existing MESOS_xxx env vars, such as MESOS_FRAMEWORK_ID and MESOS_EXECUTOR_ID, that are needed to run correctly on MESOS. This makes it impossible to run spark-submit from Chronos, or any Mesos task, because the resulting driver will inherit the wrong MESOS_xxx and fail right away.
A Mesos issue reporting the same behavior: https://github.com/mesos/chronos/issues/707