The paper proposes an adaptive relation decomposition algorithm for conjunctive retrieval queries and provides some hard experimental data for its performance analysis. Our algorithm performs horizontal relation partitioning first, followed by vertical partitioning of every horizontal cut created in the first step. Since the problem of the optimal relation decomposition is NP-hard, the proposed algorithm, as based on heuristics, is not optimal. Defining the relative savings as the ratio between the cost of answering a set of queries in the relation without decomposition and the cost of answering the same set of queries in the decomposed relation, the algorithm gives significant relative savings (250–300%) for realistic test relations and query patterns. These savings can be obtained only for an application with a query pattern that is stable over a time long enough to sense the pattern and propose a relation decomposition, to actually decompose the relation (or modify its decomposition), and to pay back for the decomposition cost. A surprising result is that the relative savings do not decrease even when the pages (blocks) on which a relation is stored are small.