|
|
|
Even though the development of sophisticated join processing techniques started in the early Eighties, there are still substantial improvements possible. The need for further investigating join processing techniques is a result from the ever larger growing database volumes (i.e., in data warehouse applications) that need to be processed.
The paper by Li and Ross describes two new join algorithms that are based on join indices. A join index materializes the join between two relations, say R and S, by storing all the RID-pairs of those tuples r,s that satisfy the join predicate. One of the algorithm is sort based, the other is partition based. Both algorithms have in common, that they store the join rsult on disk such that two files are crated: one with R tuples and one with S tuples which need to be merged to obtain the join result. Therefore, their algorithm achieves ist performance gain mostly in situations where the join needs to be materialized (e.g., in precomputing join views) and less in situations where the join result is further processed (then the tuples have to be re-read from disk because their algorithm does not pipeline tuples to the next operator). It appears to me, that the techniques developed in this paper for join index joins can be adapted for evaluating functional joins along logical object identifiers. The mapping structure for transforming a logical OID into a physical address can then be viewed as the join index.
Copyright © 1999 by the author(s). Review published with permission.