The goal is to reduce the number of value sent to a single post-GBK worker. If combiner lifting happens, each bundle will sends a single value per sub-key, causing an N-fold blowup in shuffle data and N reducers with the same amount of data to consume as the single reducer in the non-fanout case.
Imported from Jira BEAM-4565. Original Jira may contain additional context.
Reported by: robertwb.