angel

Why we included this project

Angel targets the kind of model training where the feature space grows so large that a single machine cannot hold the parameters. It is a distributed parameter server that spreads model state across multiple nodes, and Tencent has run it on exactly those industrial-scale workloads. The repo ships ready-to-run implementations of logistic regression, GBDT, factorization machines, and graph methods such as PageRank and word2vec, so you are not assembling everything from scratch. Most teams will get the most value from Spark on Angel, which keeps your familiar DataFrame pipelines while moving model weights down to the parameter-server layer. Written in Java and Scala and running on Yarn.