When using DataflowRunner as an underlying runner and GCS buckets as cache dirs, the ib.show() and ib.collect() misbehaves because the cache manager does not wait for cache file to be written while DataflowRunner running on Dataflow service takes a long time to complete the write tasks.
Imported from Jira BEAM-11692. Original Jira may contain additional context.
Reported by: ningk.