Hash, Sort-Merge, Broadcast: How Distributed Joins Work

This article is part 9 of a series on query engine design, covering how distributed databases handle joins across nodes. It details three main strategies: shuffle join (redistributing both tables), broadcast join (copying the small table to all nodes), and co-located join (no data movement). It also compares local join algorithms like hash join and sort-merge join, and discusses how optimizers choose between them. The key insight is that distributed joins are network-bound, unlike single-node CPU-bound joins. Aimed at developers and engineers working with distributed SQL engines like Spark, Snowflake, and BigQuery.