Distributed filesystem metadata

Design the metadata service for a distributed filesystem (HDFS/GFS-style): the system stores huge files as fixed-size blocks replicated across thousands of data nodes, but a single namespace must track the directory tree, file to block mappings, and block to replica locations. Clients do path lookups, opens, and renames; the metadata layer must be fast, consistent, and not become a single point of failure. How do you design it?