LightLDA is a distributed system for large scale topic modeling. It implements a distributed sampler that enables very large data sizes and models. LightLDA improves sampling throughput and convergence speed via a fast O(1) metropolis-Hastings algorithm, and allows small cluster to tackle very large data and model sizes through model scheduling and data parallelism architecture. LightLDA is implemented with C++ for performance consideration. We have sucessfully trained big topic models (with trillions of parameters) on big data (Top 10% PageRank values of Bing indexed page, containing billions of documents) in Microsoft.