[BEAM-9877] Eager size estimation of large group-by-key iterables cause expensive / duplicate reads

• Type: Bug
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Fix Version/s: 2.21.0
• Component/s: sdk-java-core
• Labels: Done
• Assignee: Tudor Marian
• Reporter: Tudor Marian
• Created: 04/May/20 17:45
• Updated: 06/Oct/20 21:55
• Resolved: 04/May/20 20:20

Description

BatchGroupAlsoByWindowViaIteratorsFn WindowReiterator can cause very expensive duplicate reading of data from a (Co-)GroupByKey for hot keys with many values due to PCollection size estimation.  Instead, it should perform lazy estimation like GroupingShuffleReader and GroupingShuffleEntryInterator (or perform no estimation at all).

Issue Links

• links to: GitHub Pull Request #11598
• links to: GitHub Pull Request #11601