September 19 – 24 , 2010, Dagstuhl Seminar 10381

Robust Query Processing


Goetz Graefe (HP Labs – Madison, US)
Arnd Christian König (Microsoft Corporation – Redmond, US)
Harumi Anne Kuno (HP Labs – Palo Alto, US)
Volker Markl (TU Berlin, DE)
Kai-Uwe Sattler (TU Ilmenau, DE)

For support, please contact

Dagstuhl Service Team


Dagstuhl Seminar Proceedings DROPS
List of Participants
Dagstuhl's Impact: Documents available


In the context of data management, robustness is usually associated with recovery from failure, redundancy, disaster preparedness, etc. Robust query processing, on the other hand, is about robustness of performance and scalability. It is more than progress reporting or predictability. A system that predictably fails or obviously performs poorly is somewhat more useful than an unpredictable one, but it is not robust. This is comparable to an automobile that only starts in dry weather: it is predictable but not nearly as useful or robust as a car that starts in any weather.

Robust query processing performance has been a known problem for a long time. It also seems common to most or all database management systems and most or all installations. All experienced database administrators know of sudden disruptions of data center processing due to database queries performing poorly, including queries that had performed flawlessly or at least acceptably for days or weeks.

We believe that a fundamental cause of lack of robustness is that the various stages of database query processing are performed by loosely coupled system components developed, maintained, and studied by largely disjoint cliques of developers and researchers. Only a handful of researchers have established expertise in more than one, or possibly two, areas of query processing. In many industrial database development groups, the query optimizer and executor teams report to different management chains.

Some techniques are meant to alleviate problems of poor performance, e.g., automatic index tuning or statistics gathered and refreshed on-demand. However, they sometime exacerbate the problem. For example, insertion of a few new rows into a large table might trigger an automatic update of statistics, which uses a different sample than the prior one, which leads to slightly different histograms, which results in slightly different cardinality or cost estimates, which leads to an entirely different query execution plan, which might actually perform much worse than the prior one due to estimation errors. Such occasional "automatic disasters" are difficult to spot and usually require lengthy and expensive root cause analysis, often at an inconvenient time.

A frequent cause of unpredictable performance is that compile-time query optimization is liable to suffer from inaccuracy in cardinality estimation or in cost calculations. Such errors are common in queries with dozens of tables or views, typically generated by software for business intelligence or for mapping objects to relational databases. Estimation errors do not necessarily lead to poor query execution plans, but they do so often and at unpredictable times.

Other sources for surprising query performance are widely fluctuating workloads, conflicts in concurrency control, changes in physical database design, rigid resource management such as a fixed-size in-memory workspace for sorting, and, of course, automatic tuning of physical database design or of server parameters such as memory allocation for specific purposes such as sorting or index creation.

Numerous approaches and partial solutions have been proposed over the decades, from automatic index tuning, automatic database statistics, self-correcting cardinality estimation in query optimization, dynamic resource management, adaptive workload management, and many more. Many of them are indeed practical and promising, but there is no way of comparing the value of competing techniques (and they all compete at least for implementation engineers!) until a useful metric for query processing robustness has been defined. Thus, defining robustness as well as a metric for it is a crucial step towards making progress.

Such a metric can serve multiple purposes. The most mundane purpose might be regression testing, i.e., to ensure that progress, once achieved in a code base, is not lost in subsequent maintenance or improvement of seemingly unrelated code or functionality. The most public purpose might be to compare competing software packages in terms of their robustness in query processing performance and scalability as a complement to existing benchmarks that measure raw performance and scalability without regard to robustness.

Dagstuhl Seminar Series


  • Data Bases / Information Retrieval
  • Data Structures / Algorithms / Complexity
  • Optimization / Scheduling
  • Interdisciplinary


  • Robust query processing
  • Adaptive query optimization
  • Query execution


In the series Dagstuhl Reports each Dagstuhl Seminar and Dagstuhl Perspectives Workshop is documented. The seminar organizers, in cooperation with the collector, prepare a report that includes contributions from the participants' talks together with a summary of the seminar.


Download overview leaflet (PDF).

Dagstuhl's Impact

Please inform us when a publication was published as a result from your seminar. These publications are listed in the category Dagstuhl's Impact and are presented on a special shelf on the ground floor of the library.


Furthermore, a comprehensive peer-reviewed collection of research papers can be published in the series Dagstuhl Follow-Ups.