It is a misconception that partitioning search keys improves the reliability of a name search. Partitioning will always result in some loss of reliability. However, all name search systems are susceptible to a conflict between performance and reliability. When extreme volumes of data are to be searched, and performance is more critical than reliability, there is a case for partitioning the keys.

The choice of what data to partition with also creates a conflict between quality and performance. An attribute which achieves the performance objective, but is not measurably reliable, is not helpful. An attribute which is measurably reliable, but does not meet the performance objective, is also not helpful.

For some systems a year of birth may be a good partition, but no good if the error rate in birth dates is high. For other systems a state code may be a good partition, but no good if there is a high rate of movement between states, or a lack of truth in the state codes.

The need for partitioning should be empirically derived (as a result of tests on real data, in real volumes, in a production-like environment) and not decided upon theoretically. If partitioning is used, when null or suspicious values of the partitioning attribute are encountered, these must be added to a common partition which is searched whenever a specific partition is searched. Also when nulls or errors are found in the search data’s partitioning attribute then all partitions must be searched.

A strong search system will allow searches across all partitions, even if this is not the default search.