Database Reference
In-Depth Information
By using version management instead of record-wise locking techniques
(which becomes possible largely because of the typically much smaller
number of concurrent updates), each query would see a consistent
database state in which no locks, or very few locks, ever occur. Also,
version management is what complex analytics users need to be able to
keep track of their many different analysis paths across the database.
It is possible (although probably often not necessary) to index every attribute
since searches greatly dominate over updates and because adding an in-
dex to an attribute requires only that attribute to be read, not the entire
row.
Data compression is likely to be profitable because the data belonging to one
attribute are highly likely to be homogeneous and even auto-correlated.
Adding or deleting attributes of a table is likely to be cheap since only rel-
evant data would be accessed. Updates of an attribute are likely to be
relatively cheap because no irrelevant attribute values need to be read
or written.
Similar observations were made much earlier,
3
but were for a long time con-
sidered irrelevant in mainstream database research. In fact, these alternative
design principles have only comparatively recently received serious attention.
This renewed interest is at least partly due to the fact that today, many very
large databases are actually used for data warehousing, decision support, or
business or security intelligence applications, areas where similar character-
istics apply as those claimed above. In Svensson,
3
the following additional
observations were made:
In scientific data analysis, the number of simultaneous users is typically
much smaller than in large-scale commercial applications; but on the
other hand the users tend to put more complex, usually unanticipated,
queries to the system.
A more automatic response to complex user requests is required from the sys-
tem components, since in scientific applications no systems or database
specialists are usually available.
A system should be transportable to many computer models, including
medium-sized computers, because in scientific data analysis applications
many users prefer to work with an in-house computer dedicated to the
acquisition and analysis of data.
One of the first systems to be developed based on the above principles was
the system
Cantor
.
4
-
7
The Cantor project pioneered the analysis and coor-
dinated application of many of the above-mentioned techniques and concepts
in relational systems.
