The Google File System
Google described a file system that stores petabytes across thousands of cheap machines and treats their failure as normal rather than as an incident.
Why it matters
Storage stopped requiring reliable hardware: reliability moved into software and redundancy.
Each file is split into chunks and stored in several copies; a machine failing is not an event requiring intervention. That changed the economics: many cheap machines instead of expensive servers. The consequence for machine learning is direct, because training at web scale first requires the ability to store that scale. Hadoop reproduced the system in open source.