MapReduce
Google's programming model for running batch jobs across thousands of machines without the developer thinking about distribution.
Two functions: map turns each input record into key–value pairs, reduce aggregates all values per key. Between them the framework performs a shuffle that groups by key. Fault tolerance comes from restarting individual tasks.
The paper (Dean & Ghemawat, 2004) spawned Hadoop and with it the whole big data era. The model is largely historical now — Spark does the same thing in memory and is orders of magnitude faster — but the shuffle step lives on inside every distributed query engine.