Scaling the EOS namespace

EOS is the distributed storage system being developed at CERN with the aim of fulfilling a wide range of data storage needs, ranging from physics data to user home directories. Being in production since 2011, EOS currently manages around 224 petabytes of disk space and 1.4 billion files across several instances. Even though individual EOS instances routinely manage hundreds of disk servers, users access the contents through a single, unified namespace which is exposed by the head node (MGM), and contains the metadata of all files stored on that instance. The legacy implementation keeps the entire namespace in-memory. Modifications are appended to a persistent, on-disk changelog; this way, the in-memory contents can be reconstructed after every reboot by replaying the changelog. While this solution has proven reliable and effective, we are quickly approaching the limits of its scalability. In this paper, we present our new implementation which is currently in testing. We have designed and implemented QuarkDB, a highly available, strongly consistent distributed database which exposes a subset of the redis command set, and serves as the namespace storage backend. Using this design, the MGM now acts as a stateless write-through cache, with all metadata persisted in QuarkDB. Scalability is achieved by having multiple MGMs, each assigned to a subtree of the namespace, with clients being automatically redirected to the appropriate one.