Design considerations

Design considerations mk Thu, 02/20/2014 - 03:53

Redesign of the MonetDB software stack was driven by the need to reduce the effort to extend the system into novel directions and to reduce the Total Execution Cost (TEC). The TEC is what an end-user or application program will notice and is composed of several cost factors:

A) API message handling
P) Parsing and semantic analysis
O) Optimization and plan generation
D) Data access to the persistent store
E) Execution of the query terms
R) Result delivery to the application
 
Choosing an architecture for processing database operations presupposes an intuition on how the cost will be distributed. In an OLTP setting you expect most of the cost to be in (P,O), while in OLAP it will be (D,E,R). In a distributed setting the components (O,D,E) are dominant. Web-applications would focus on (A,E,R).
 
Such a simple characterization ignores the widespread differences that can be experienced at each level. To illustrate, in D) and R) it makes a big difference whether the data is already in the cache or still on disk. With E) it makes a big difference whether you are comparing two integers, evaluationing a mathematical function, e.g., Gaussian, or evalauting a regular expression on a string. As a result, intense optimization in one area may become completely invisible due to being overshadowed by other cost factors.

The Version 5 infrastructure is designed to ease addressing each of these cost factors in a well-defined way, while retaining the flexibility to combine the components needed for a particular situation. It results in an architecture where you assemble the components for a particular application domain and hardware platform.

The primary interface to the database kernel is still based on the exchange of text in the form of queries and simply formatted results. This interface is designed for ease of interpretation and versatility, and is flexible to accommodate system debugging and application tool development. Although a textual interface potentially leads to a performance degradation, our experience with earlier system versions showed that the overhead can be kept within acceptable bounds. Moreover, a textual interface reduces the programming effort otherwise needed to develop test and application programs.