• No se han encontrado resultados

MARCO TEÓRICO

In document FACULTAD DE CIENCIAS EMPRESARIALES (página 11-18)

Apart from the containment problem, a multi-view based approach also suffers from the issue of determining what part of the views should be materialised to avoid duplication and to expedite the view adaptation process. Existing selection mechanisms fall into three parts which we refer to them as plan-based selection, full materialisation selection and partial materialisation selection, respectively.

Many research efforts have been proposed for cost-based plan selection in relational database systems. In [GPSH02], the authors provide a cluster based selection plan, where queries are grouped based on their corresponding similarities. In their approach, they store a database of plans and attempt to assign one of these plans to the new query with the expectation that the selected plan would be the same as the plan generated by the optimiser. Only in the event that no suitable assignment can be found, is the optimisation process actually carried out and the newly generated plan is added to the plan database for future use. In [GCV09], the authors present a cost-based optimisation and execution framework. Their selection algorithm is based on a navigation-focused XPath algebra with novel operators and a com-prehensive set of rewriting rules. By evaluating the costs of different algebraic representa-tions after applying rewriting rules, their approach can select best plans for query execution.

However, while these approaches provide an example of selecting the appropriate view to answer a query, they do not solve the problem of selecting views for materialisation.

On the other hand, there has been considerable research on full materialisation selection for relational database systems and this can be broadly categorised in two ways: centralised approaches [GZ08, KGJ10] and distributed approaches [BL03, YGYL05, CBHB09]. The former chooses materialised views in centralised scenarios, where storage is considered

to be the limiting factor while the later chooses materialised views in distributed environ-ments, where the primary factor for concern is network communication costs. However, all approaches focuses on selecting entire views for materialisation and do not consider the potential shareability between views.

The partial materialisation mechanism discussed earlier in this chapter [Bel04], tries to balance the query processing cost and view maintenance cost. In their approach, each OR-Node (fragment) in the view graph is associated to a level, where the top OR-node has the level value 1. When an OR-node at level 1 is materialised, the query processing cost is low as the result of that OR-node is the result of the view, but the maintenance cost is high as the entire result is stored. However, if the OR-node at any other level is materialised, then the query processing cost is relatively high, but the maintenance cost is low. The purpose of their approach is to find an intermediary level for materialisation which balances the cost of query processing cost and view maintenance. The selection algorithm is based on two measures: local benefit and global benefit. A view is considered locally beneficial if its materialisation significantly reduces the query processing cost without significantly increasing the view maintenance. The global benefit is the measure of the importance of a fragment to all views in the view graph. Their selection is a two-step process, where the first step is to select a set of fragments within each query based on the local benefit and the second part filters fragments selected in the first part according to their global benefit to the entire view graph. In their approach, a suboptimal decision may be made when materialised views are dissimilar to each other, leading to potentially poor performance during view adaptation.

Summary and Issues. When there are a large number of views, merely calculating the benefit of each fragment to all views is not an appropriate solution. A more desirable ap-proach would take user preferences into account. Suppose the benefit of a fragment is high, but it is only shared by views that are rarely required by the users, therefore, there is no rea-son to materialise such a fragment. Besides, views should only compare to other existing views that are similar to them as this provides a more accurate measure of the benefit that a fragment contributing to existing views. As we will shown in our experiments, view adap-tation performs better by attempting to optimise the selection fragments for materialisation.

2.10 Summary

In this chapter, different view adaptation approaches were examined and analysed. Early work failed to “adapt” materialised data, and instead simply hid the old materialisation from the new view, or referenced the old materialisation when attribute type is changed.

The centralised approach discussed in §2.3 provided a more comprehensive view adapta-tion mechanism that could be used to handle different changes. However, they did not take network communication costs into account, a key feature in §2.4. This work utilised two auxiliary attributes to save the cost of transferring redundant data over the network. To max-imise reuse, the ExpressionTree approach divided views into different nodes and analysed when old materialisation can be reused and when intermediate results are required.

The MultiView approach was shown to outperform the single view based approaches by sharing materialised data between views to improve not only the view adaptation perfor-mance, but also the reusability of materialised data. Additionally, the cost of maintaining materialised data is also reduced. However, existing XML view adaptation does not em-ploy a multi-view approach, with view adaptation algorithms based on single views, and changes are very restricted. Thus, views must often be recomputed with data retrieved from data sources. Strongly motivated by this literature review, the next chapter will provide an outline of a fragment based adaptation system for XML repositories.

Chapter 3

In document FACULTAD DE CIENCIAS EMPRESARIALES (página 11-18)

Documento similar