The basic idea is simple: by pre-organizing data in a special way, your queries can run faster.
This idea has been widely-used in the various databases, including traditional RDBMS and modern distributed SQL-on-Hadoop systems. The examples are indexing, columnar format, compression, etc.
Our system, VerdictDB, uses a new orthogonal approach, which we call scrambling (more details follow shortly). Since our approach is orthogonal to the existing approaches (e.g., columnar formats and compressions), we can combine scrambling with other approaches as well; in fact, VerdictDB also uses columnar format and compression (as well as scrambling) for maximum possible speedups whenever possible.
VerdictDB mainly focuses on speeding up aggregate queries, i.e., the queries including aggregate functions, such as
avg. Unlike lookup-style queries, these queries typically require extensive data scans; thus, query latencies can be very slow especially when the size of data is large.
Based on these simple aggregate queries, more advanced analytics can also be performed. We describe them on a separate article.
Then, what is scrambling and how do queries run faster with it? We present conceptual ideas below. VerdictDB's deployment, user interface, and architecture are discussed on this page. More details about internal query processing logic is described on this page.
A scramble is a regular database table that consists of randomly shuffle tuples (of an original table) with an extra, augmented column named
verdidctdbblock. Physically, a set of tuples associated with the same
verdidctdbblock are clustered (using
partition supported by most modern databases).
VerdictDB creates a scramble when the user issues a create-scramble query.
How do queries run faster¶
To quickly produce the answers to aggregate queries, VerdictDB relies on a well-known statistical property called the law of large numbers. The law of large numbers indicates that many commonly-used statistics (e.g., mean) can be precisely estimated using a sample, a randomly chosen subset of the original data. Observe that a set of tuples associated with the
verdictdbblock in a scramble amounts to a sample; thus, by processing one or just a few blocks, we can produce very accurate estimates of the exact answer, i.e., the answer you would get if the entire data is processed.
In some situations, however, only the exact answers are needed, and you may think VerdictDB's approximate answers not directly usable. However, even in that case, VerdictDB still offers benefits with its streaming query engine. The streaming query engine progressively updates its estimates while the exact answer is being computed. With this mechanism, you know what your exact answers will be like even before all data is processed, while the more accurate answers are continuously computed in the background. If the estimates are much different from what you originally expected (e.g., wrong filtering conditions, wrong tables or columns specified, etc.), you can simply stop the execution in the middle. Find more details about the streaming query engine on this page.
More technical details can be found in our research paper.