MapReduce
Dean and Ghemawat reduced distributed computation to two functions and hid all the complexity of failures and work distribution inside the system.
Why it matters
Processing data across a thousand machines became available to a programmer who knows nothing about distributed systems.
The programmer writes only the transformation of each record and the reduction of results; splitting, restarting after failures and balancing are the system's business. The model's restrictiveness proved its advantage: a very large share of real tasks fits inside it. Two years later Hadoop gave an open implementation, and the scheme became an industry standard for a decade.