MapReduce
Google describes how it processes the whole web on thousands of cheap machines with two functions; the infrastructure for training on internet-scale data becomes ordinary.
what had to happen · 0 events back to 1943
A root. Nothing in the archive precedes it.
Jeffrey Dean and Sanjay Ghemawat's paper at the OSDI conference in December 2004 described the system Google used to build its index. A programmer writes a map function, which turns each input record into key–value pairs, and a reduce function, which combines the values for each key; the framework does everything else, splitting the data across thousands of commodity machines, restarting the ones that fail, and moving the results. Petabytes became a routine unit.
The paper is on this timeline because scale needed plumbing. The deep-learning era is usually told as a story of algorithms and GPUs, but the data those GPUs consumed, the crawled web, the billions of images, the scraped text of the 2020s, was collected, cleaned and filtered with MapReduce and its descendants. Hadoop, the open-source reimplementation, made the same capability available to anyone from 2006, and the Common Crawl corpus that trained GPT-3 was produced with it.
Dean went on to co-found Google Brain in 2011 and to lead the engineering of TensorFlow and the TPU. The people who built the systems for indexing the web were, a few years later, the people who built the systems for learning from it.
what it led to · 0 events downstream
A leaf, for now. Nothing in the archive has built on it yet.