Function-based indexing for object-oriented databases

Object-oriented databases have been developed to allow applications to store persistent data in the form of objects. As in conventional databases, object-oriented databases need to support queries that allow associative access to data by describing a property of the data of interest. In particular, they should support queries over user-de ned sets based on user-de ned functions that compute the property of interest. To make such function-based queries run faster, we must support function-based indexes where the keys are these computed properties. Function-based indexes are harder to maintain than indexes in conventional databases because mutations that a ect keys can be done without accessing an element of the indexed set. This dissertation presents a new function-based indexing scheme for object-oriented databases. The indexes are maintained by recording registration information for objects that indicate when index entries are a ected and having operations that modify objects check for registrations to determine if an index needs to be updated. Our approach uses a combination of declared information that indicates how object methods a ect each other and runtime information gathered during key computation, so that we register only the objects whose modi cations can a ect an index and only the operations that might a ect an index entry check for registration information. Recomputing index entries only when there is a modi cation that might a ect an index comes at the cost of extra space for the registration information and extra time to check registration information during modi cations. To quantify these costs, we simulated three implementations of our scheme on various performance benchmarks and analyzed several others. We show that function-based indexes are very useful in object-oriented databases and do not adversely a ect system performance, but they do have a high space cost. This work is being done in the context of the Thor distributed, object-oriented database system[42, 43]. Thor has features that complicate index use and maintenance. It has a clientserver architecture with caching at both clients and servers. It also uses an optimistic concurrency control scheme. This dissertation presents a design for integrating index use and maintenance into Thor. We discuss the impact of adding indexes on Thor's system architecture and the impact that Thor's system architecture has on the performance of our indexing scheme, and we present predicate validation, a new optimistic concurrency control scheme. Thesis supervisor: Barbara H. Liskov Title: NEC Professor of Software Science and Engineering