[SPARK-27659] Allow PySpark toLocalIterator to prefetch data

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 3.0.0
• Fix Version/s: 3.0.0
• Component/s: PySpark
• Assignee: Holden Karau
• Reporter: Bryan Cutler
• Created: 08/May/19 17:39
• Updated: 04/Dec/19 01:58
• Resolved: 21/Sep/19 01:20

Description

In SPARK-23961, data was no longer prefetched so that the local iterator could close cleanly in the case of not consuming all of the data. If the user intends to iterate over all elements, then prefetching data could bring back any lost performance. We would need to run some tests to see if the performance is worth it due to additional complexity and extra usage of memory. The option to prefetch could adding as a user conf, and it's possible this could improve the Scala toLocalIterator also.

Issue Links

• is duplicated by: SPARK-27025 Speed up toLocalIterator
• is related to: SPARK-23961 pyspark toLocalIterator throws an exception
• links to: GitHub Pull Request #25515