[SPARK-42689] Allow ShuffleDriverComponent to declare if shuffle data is reliably stored

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.5.0
• Fix Version/s: 3.5.0
• Component/s: Spark Core
• Assignee: Mridul Muralidharan
• Reporter: Mridul Muralidharan
• Created: 06/Mar/23 19:24
• Updated: 09/Mar/23 07:10
• Resolved: 09/Mar/23 07:10

Description

Currently, if there is an executor node loss, we assume the shuffle data on that node is also lost. This is not necessarily the case if there is a shuffle component managing the shuffle data and reliably maintaining it (for example, in distributed filesystem or in a disaggregated shuffle cluster).

Downstream projects have patches to Apache Spark in order to workaround this issue, for example Apache Celeborn has this.

Issue Links

• links to: [Github] Pull Request #40307 (mridulm)
• links to: [Github] Pull Request #40307 (mridulm)