We present an optimization method and algorithm designed for three objectives: physical data independence, semantic optimization, and generalized tableau minimization. The method relies on generalized forms of chase and "backchase" with constraints (dependencies). By using dictionaries (finite functions) in physical schemas we can capture with constraints useful access structures such as indexes, materialized views, source capabilities, access support relations, gmaps, etc. The search space for query plans is de ned and enumerated in a novel manner: the chase phase rewrites the original query into a "universal" plan that integrates all the access structures and alternative pathways that are allowed by applicable constraints. Then, the backchase phase produces optimal plans by eliminating various combinations of redundancies, again according to constraints. This method is applicable (sound) to a large class of queries, physical access structures, and semantic constraints. We prove that it is in fact complete for "path-conjunctive" queries and views with complex objects, classes and dictionaries, going beyond previous theoretical work on processing queries using materialized views.

Traditionally, query optimizers assume a direct mapping from the logical entities modeling the data (e.g. relations) and the physical entities storing the data (e.g. indexes), each physical entity corresponding precisely to one logical entity. This assumption is no longer true in non-traditional applications (object-oriented and semi-structured databases, data integration), which often exhibit a mismatch between the logical view and the actual storage of data. In addition, there is an increased amount of redundancy, even at the logical level, that can greatly enhance optimization opportunities, if exploited. To deal with all this, we propose a novel architecture for query optimization, in which physical optimization is leveraged at the level of query rewriting. As a consequence, the other important aspect of query optimization, semantic optimization (that takes advantage of the redundancy at the logical level), can be naturally incorporated. The optimizer can then make global decisions based on both semantic and physical knowledge, leading to plans of higher quality than those obtainable by a traditional two-level approach. The main idea

Abstract. Previous research on materialized views has primarily been in the context of flat relational databases--materialized views defined in terms of one or more flat relations. This paper discusses a broader class of view definitions--materialized views defined over a nested data model such as the nested relational model or an object-oriented data model. An attribute of a tuple deriving the view can be a reference (i.e., a pointer) to a nested relation, with arbitrary levels of nesting possible. The extended capability of this nested data model, together with materialized views, simplifies data modeling and gives more flexibility. Simple extensions of standard view maintenance techniques to the nested model would do too much work for maintenance: a change in a nested set would re-process the entire nested set, not just the changed parts. We show how existing incremental maintenance algorithms can be extended to maintain the views without performing this additional work. We describe the implementation of these techniques in the SWORD in-terface to the Ode database system. The implementation is based on the representation of nested structures by classes and the use of an SQL-like language to define materialized views. We outline the data structures and algorithms used in the implementation and examine performance. This is one of the first pieces of work to explore the applicability of materialized views over complex objects. 1

This dissertation presents a new parallel object-oriented database system implementation and architecture. The system, parallel ADAMS, we have implemented as appropriate to large-scale scientific database applications, where the retrieval of complex data from very large collections is a primary operation. Aside from being a parallel implementation, parallel ADAMS differs from typical OODBMSs in three significant ways: (1) it employs the decomposed storage model, rather than contiguous object storage, (2) it is based on a query server architecture, and (3) it employs a shared nothing distributed architecture. Parallel ADAMS sets are partitioned by oid. In the dissertation, we demonstrate that set operators, and therefore logical query connectives, can be performed in a completely data parallel fashion. More complex queries involving implicit joins require inter-processor communication which is minimized in our implementation. The implementation runs on general purpose hardware. Results ...

I am indebted to Val Tannen, my advisor. This dissertation would not have been possible without his invaluable ideas, support, and advice. His deep insight into both elds of databases and programming languages, as well as his crucial advice in the experimental issues, made direct contributions to this work. I have learned a great deal from him: where to look for ideas, how to solve problems, how to present them. I also have learned from him that most rewarding results are obtained by taking the path of investigating the foundations, even though this may not be the shortest path. Iamvery grateful to Peter Buneman and Susan Davidson, for their help, support, advice, suggestions and encouragement. I am thankful to Alin Deutsch who was always a source of inspiration. Many ideas coming from fruitful discussions with him have entered in this dissertation. I would like togivespecial thanks to Arnaud Sahuguet who always provided me with valuable suggestions and comments, as well as with expert help regarding system and implementation issues. Iwould like to thank the members of my thesis committee for their insightful comments, suggestions, questions and criticism: Peter Buneman, Susan Davidson, Daniela Florescu, Jean Gallier and Alon Levy. This work has also bene ted greatly from the discussions at the weekly Penn database seminar, whose members include (or included during my veyears of PhD study) Peter