Understanding the shuffle bottleneck

Your manager at Global Retail Analytics asks you to enrich the online_retail dataset (over 100,000 rows) with region data from country_lookup (44 rows). You write a standard left join, but the query takes several minutes. A senior engineer checks your .explain() output, spots ShuffleExchange nodes, and tells you: "The shuffle is your bottleneck."

What does the senior engineer mean by "the shuffle"?