“The Google File System” by Sanjay Ghemawat, Howard Gobioff, and Shun-Tak Leung was presented at the 19th ACM Symposium on Operating Systems Principles (SOSP) in 2003. The paper describes a scalable distributed file system that Google built to support its large, data-intensive applications, and which the authors say was already in production use across Google’s infrastructure at the time of writing.
The central design decision was to assume that hardware failure is normal rather than exceptional. Because the system runs on inexpensive commodity machines that fail constantly, the file system was designed for continuous monitoring, automatic recovery, and replication built into its core. Files are split into large fixed-size chunks, each replicated across multiple machines, so the loss of any single machine does not lose data.
The architecture uses a single master server that manages metadata and many chunkservers that store the actual data. The authors optimized for the access patterns they actually observed at Google: files are huge, writes are mostly appends rather than overwrites, and most reads stream large sequential portions of data. These assumptions let them relax the guarantees a general-purpose file system would need and instead tune aggressively for throughput at scale.
The Google File System became one of the most influential systems papers of its era. Its design directly inspired the open-source Hadoop Distributed File System (HDFS), and together with the MapReduce and Bigtable papers it laid the foundation for the modern big-data ecosystem.