Dan Suciu
University of Washington, WA·Computer Science
Before a computer runs a big data-processing job, it helps enormously to know roughly how much data that job will produce. Database systems use such estimates to decide how much memory and compute to reserve and, crucially, to pick among many possible ways of executing the same query.
- Fields
Computer Science
Mathematics & Statistics
- Topics
cardinality estimation
query optimization
information theory
- Grants
- 2 known grants, through Sep 2028
What they work on
- Compute upper bounds on query output size that are guaranteed never to be exceeded, deriving the tightest possible bound from information theory given statistics a database already stores, such as table cardinalities, distinct-value counts, and maximum degrees.
- Develop efficient algorithms for evaluating these information-theoretic bounds, and test whether Lp-sketches, compact summaries of how unevenly column values repeat, can collect and incrementally maintain the needed statistics offline.
- Extend pessimistic cardinality estimation to group-by queries, where conventional estimation methods are known to perform very poorly, including refinement by bucketing the input data as histograms do today.
- Enable aggregates such as summation and counting inside Datalog, a recursive query language, by generalizing relations to relations over a semiring so recursion and aggregation can be interleaved freely.
- Characterize the complexity, optimization, and evaluation strategies of semiring-based Datalog, testing whether standard SQL optimizations carry over to the recursive, aggregating setting.
More about the research
Bad estimates waste energy or cause jobs to fail. Work in this lab attacks two sides of this problem: making query-size predictions that come with mathematical guarantees, and extending database query languages so that modern data-science tasks, which require repetition and arithmetic, can run inside the database engine itself.
Why it matters
Query size estimates drive how database systems allocate memory and choose execution plans, so guaranteed upper bounds could cut wasted compute and energy and prevent failed jobs on large workloads. Letting a relational database express iteration together with sums and counts could put machine-learning-style analytics within reach of data scientists using familiar database tools, rather than requiring specialized machine-learning expertise.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Andrew Pavlo
Carnegie Mellon University, PA·Computer Science
Develop automatic code transformations that speed up user-defined functions (UDFs) inside SQL queries without requiring changes to application code.
Joy Arulraj
Georgia Institute of Technology, GA·Computer Science·Early-career
Design query optimizers tailored to offline, exploratory video analytics that use accuracy-driven optimization to reduce the cost of running complex computer vision pipelines over large video datasets.
Zeyu Ding
Binghamton University, NY·Computer Science
Build automated tools that analyze an SQL database schema and help users interactively define a customizable privacy model specifying which data elements need differential-privacy protection, and which variant.
The school
University of Washington: 81 labs on PI Pages, $1.7B a year on research, 694 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.