This paper examines the feasibility of using long term (disk based) caches in very large distributed file systems (DFSs). We begin with an analysis of file access patterns in a distributed Unix workstation environment, and identify properties of use to the DFS designer. We then introduce long-term caching strategies that maintain consistency while dramatically reducing the load on file servers. We describe a number of algorithms for maintaining client caches, and present the results of a trace-driven simulation that shows how relatively small disk-based caches can be used to reduce server traffic by 60% to 90%. Finally, we outline possible mechanisms for dynamically organizing these caches into adaptive hierarchies to allow arbitrary scaling of the number of clients and the use of low-bandwidth communication networks. A small (2 or 3 level) hierarchy, coupled with smart caching techniques, has the potential to reduce traffic by an order of magnitude or more over a flat scheme.