I've been reading The Chubby Lock Service for Loosely-Coupled Distributed Systems by Mike Burrows. It describes Chubby, which is used at Google by both GFS and BigTable. Out of the whole fascinating paper I think Section 4, "Use, surprises and design errors" was far and away the best part. Some example quotes:
Chubby’s C++ client library is 7000 lines (comparable with the server), and the client protocol is delicate. To maintain the library in Java would require care and expense, while an implementation without caching would burden the Chubby servers. Thus our Java users run copies of a protocol-conversion server that exports a simple RPC protocol that corresponds closely to Chubby’s client API. Even with hindsight, it is not obvious how we might have avoided the cost of writing, running and maintaining this additional server.
Chubby was never intended to be used as a storage system for large amounts of data, and so it has no storage quotas. In hindsight, this was naïve.
Some of our remedies are heavy-handed. For example, we review the ways project teams plan to use Chubby, and deny access to the shared Chubby name space until review is satisfactory. A problem with this approach is that developers are often unable to predict how their services will be used in the future, and how use will grow. Readers will note the irony of our own failure to predict how Chubby itself would be used.