attracted the broadest interest within the group.

In addition to the listed topics, the main issues raised during the meeting included management of uncertain information, data privacy and security, e-science and other scholarly applications, human-centric interaction with data, social networks and Web 2.0, personalization and contextual-ization of query- and search-related tasks, streaming and networked data, self-tuning and adaptive systems, and the challenges raised by new hardware technologies and energy constraints. Most are captured in the following discussion, with many cutting across multiple topics.

Revisiting database engines. System R and Ingres pioneered the architecture and algorithms of relational databases; current commercial databases are still based on their designs. But many of the changes in applications and technology demand a reformation of the entire system stack for data management. Current big-market relational database systems have well-known limitations. While they provide a range of features, they have only narrow regimes in which they provide peak performance; online transaction processing (OLTP) systems are tuned for lots of small, concurrent transactional debit/credit workloads, while decision-support systems are tuned for a few read-mostly, large-join-and-aggregation workloads. Meanwhile, for many popular data-intensive tasks developed over the past decade, relational databases provide poor price/performance and have been rejected; critical scenarios include text indexing, serving Web pages, and media delivery. New workloads are emerging in the sciences, Web 2.0-style applications, and other environments where database-engine technology could prove useful but is not bundled in current database systems.

Even within traditional application domains, the database marketplace today suggests there is room for significant innovation. For example, in the analytics markets for business and science, customers can buy petabytes of storage and thousands of processors, but the dominant commercial database systems typically cannot scale that far for many workloads. Even when they can, the cost of software and

the ubiquity
of big data is
expanding the
base of users
and developers of
data-management
technologies and
will undoubtedly
shake up
the database
research field.

management relative to hardware is exorbitant. In the OLTP market, business imperatives like regulatory compliance and rapid response to changing business conditions raise the need to address data life-cycle issues (such as data provenance, schema evolution, and versioning).

Given these requirements, the commercial database market is wide open to new ideas and systems, as reflected in the recent funding climate for entrepreneurs. It is difficult to recall when there were so many startup companies developing database engines, and the challenging economy has not trimmed the field much. The market will undoubtedly consolidate over time, but things are changing fast, and it remains a good time to try radical ideas.

Some research projects have begun taking revolutionary steps in database system architecture. There are two distinct directions: broadening the useful range of applicability for multipurpose database systems (for example, to incorporate streams, text search, XML, and information integration) and radically improving performance by designing special-purpose database systems for specific domains (for example, read-mostly analytics, streams, and XML). Both directions have merit, and the overlap in their stated targets suggests they may be more synergistic than not. Special-purpose techniques (such as new storage and compression formats) may be reusable in more general-purpose systems, and general-purpose architectural components (such as extensible query optimizer frameworks) may help speed prototyping of new special-purpose systems.

Important research topics in the core database engine area include:

•Designing systems for clusters of many-core processors that exhibit limited and nonuniform access to off-chip memory;

• Exploiting remote RAM and Flash as persistent media, rather than relying solely on magnetic disk;

• Treating query optimization and physical data layout as a unified, adaptive, self-tuning task to be carried out continuously;

• Compressing and encrypting data at the storage layer, integrated with data layout and query optimization;