Hazy: A Relational Operator-based Approach to
                     Statistical Data Analysis
              Christopher Re, University of Wisconsin

There is an arms race to perform increasingly sophisticated data
analysis on increasingly varied types of data (text, audio, video,
OCR, sensor data, etc.). To cost effectively process these large
volumes of data, we need automatic techniques. But current systems
that process large volumes of data typically assume that the data have
a rigid, precise semantics, which these new data sources do not
possess. To understand variations in data's structure and meaning,
many of the state of the art approaches are statistical. Building an
ad hoc system for each new task seems to be wrongheaded as the lack of
code sharing makes each task expensive to complete. Additionally, such
an approach requires that we relearn the hard fought lessons of the
last 30 years of data management (e.g., handling concurrency or
recovery).

In this talk, I describe our system, Hazy, that augments a traditional
RDBMS (Postgres) to allow developers to specify statistical data
analysis applications. The hypothesis behind Hazy is that a large
fraction of a diverse set of statistical applications can be captured
using a small handful of primitives. To understand this hypothesis, my
group is building several applications to exercise Hazy's primitives.
I describe two of these applications:

First, I describe enhancing a text processing application with a
popular statistical technique (and one of Hazy's operators),
classification. A key technical challenge is to incrementally maintain
the output of the classification task (similar to maintaining a
materialized view). I describe Hazy's algorithm that gives an order of
magnitude performance improvement over naive approaches on several
real-world data sets.

Second, I describe a new collaboration that applies Hazy's ideas to
the problem of detecting neutrinos from the Big Bang. This work is in
collaboration with the IceCube team. IceCube operates a 1-cubic-km
block of ice located at the South Pole that functions as a neutrino
telescope.

In addition to describing our technical progress on these
applications, I describe how these different applications share a
common technical underpinning which supports Hazy's central
hypothesis.