Federated learning (FL), proposed by Google in 2016, has attracted increasing interest from academia and industry to address the data shortage issue of machine learning due to privacy concerns. The emerging FL paradigm outsources model training tasks to data owners instead of collecting data and saving them to a central repository. In an FL setting, a central server coordinates the learning rounds by dispatching the model under training to each client, which trains the model using her data. Thus, FL preserves user data privacy while satisfying machine learning needs. However, adopting a centralized server requires participants’ trust in the server. Moreover, federated learning does not address how to incentivize data owners to participate. Without proper reward/penalty systems, it is hard to ensure all clients behave and provide quality data for training. With features such as immutable distributed ledgers, smart contracts, and integral incentive mechanisms, blockchain has shown promises to be used with FL to solve the trust and incentive issues. This chapter reviews recently proposed Blockchain-based Federated Learning (BCFL) framework architectures and categorized them into three design patterns according to blockchain scaling mechanisms. We characterized each pattern’s applicability, advantages, and disadvantages from the network architecture, privacy and security, and incentive designs. These analyses provide insights for future framework design and three open research areas for further study.