3 ms·
Towards ML Engineering: A Brief History of TensorFlow Extended
- dekhn 6y agoThis paper is especially useful as it's the first official external Google publication describing Sibyl, a powerful but shut-down machine learning system used widely at Google until Tensorflow became popular. Sibyl was an interesting system for a number of reasons; it was linear and logistic regression implemented using MapReduce. It could train quickly over large data. It had many utilities for creating complicated pipelines. It also used every quirk of the google environment to obtain high performance. I worked on it for a while, and was really glad to see it retired, as the underlying design had some serious problems at scale. But it's still useful that its existence and some of the techniques it used, which have been migrated to TFX, are documented officially.