[SPARK-8987] Increase test coverage of DAGScheduler

• Type: Umbrella
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 1.0.0
• Component/s: Scheduler, Spark Core, Tests
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Andrew Or
• Created: 10/Jul/15 19:16
• Updated: 17/May/20 17:48
• Resolved: 21/May/19 04:33

Description

DAGScheduler is one of the most monstrous piece of code in Spark. Every time someone changes something there something like the following happens:

(1) Someone pings a committer
(2) The committer pings a scheduler maintainer
(3) Scheduler maintainer correctly points out bugs in the patch
(4) Author of patch fixes bug but introduces more bugs
(5) Repeat steps 3 - 4 N times
(6) Other committers / contributors jump in and start debating
(7) The patch goes stale for months

All of this happens because no one, including the committers, has high confidence that a particular change doesn't break some corner case in the scheduler. I believe one of the main issues is the lack of sufficient test coverage, which is not a luxury but a necessity for logic as complex as the DAGScheduler.

As of the writing of this JIRA, DAGScheduler has ~1500 lines, while the DAGSchedulerSuite only has ~900 lines. I would argue that the suite line count should actually be many multiples of that of the original code.

If you wish to work on this, let me know and I will assign it to you. Anyone is welcome.

Sub-Tasks

• Add end-to-end tests for the scheduling code: Resolved
• Test for fetch failure in a shared dependency for "skipped" stages: Resolved