TensorFlow GNN (TF-GNN) is a scalable library for Graph Neural Networks in TensorFlow. It is designed from the bottom up to support the kinds of rich heterogeneous graph data that occurs in today's information ecosystems. Many production models at Google use TF-GNN and it has been recently released as an open source project. In this paper, we describe the TF-GNN data model, its Keras modeling API, and relevant capabilities such as graph sampling, distributed training, and accelerator support.