Web based data integration systems run in an unpredictable and dynamic environment, which makes the traditional query processing approach inapplicable. Problems necessitating AQP (Adaptive Query Processing) techniques include: (1) statistics may be insufficient; (2) statistics may be imprecise; (3) unavailable data sources may make query results incomplete. In this paper, we present a novel data integration architecture LAD (Layered Adaptive Data integration architecture) to resolve problems mentioned above. Adopted adaptive techniques include: (1) defer making the initial plan if there are not enough high quality statistics; (2) perform re-optimizing operation if distinct deflection is detected; (3) exploit data redundancy to deal with unavailable data sources. Experiment results show that LAD provides satisfactory adaptability and efficiency in the face of uncertainty and dynamics.