When data mining and knowledge discovery techniques must be used to analyze large amounts of data, high-performance parallel and distributed computers can help to provide better computational performance and, as a consequence, deeper and more meaningful results. Recently grids, composed of large-scale, geographically distributed platforms working together, have emerged as effective architectures for high-performance decentralized computation. It is natural to consider grids as tools for distributed data-intensive applications such as data mining, but the underlying patterns of computation and data movement in such applications are different from those of more conventional high-performance computation. These differences require a different kind of grid, or at least a grid with significantly different emphases. This paper discusses the main issues, requirements, and design approaches for the implementation of grid-based knowledge discovery systems. Furthermore, some prospects and promising research directions in datacentric and knowledge-discovery oriented grids are outlined.