Conclusions

  We conclude our discussion of uniform partitioning strategies considering two practical aspects:

• It is doubtful if there is much difference between the lifespan, the range and the startpoints' span of temporal relations that are found in practice. Especially the significantly higher effort to determine the range from an IP-table in the algorithm of figure 9.4 will probably not pay off because many temporal relations have ranges that have only a few or no gaps. But it is the latter from which the benefits arise so the uniform range partitioning strategy is unlikely to provide a performance advantage that justifies the increased partitioning effort.

  In contrast to that, there is the uniform startpoints' span strategy that involves no computational disadvantage but offers a benefit in comparison to the uniform lifespan partitioning approach. The extent of this benefit is likely to be small if the lifespan is long and / or intervals are short. Both facts suggest that the startpoints' span is almost identical with the lifespan. However, the benefit might be marginal in other cases.

• It is obvious that the uniform strategies are liable to perform badly in the case of skewed , i.e. non-uniformly distributed, data. There is hardly any control over the fragments' loads and thus over the load balance . There are, however, many examples for temporal relations that have periodically repeated patterns of tuple timestamps : imagine a relation logging the starting and ending times of calls made by customers of a telephone company. The distribution of phone calls over a daytime period will vary significantly with many calls during business hours and few in the early morning and late evening. Considering a long time period, however, a daily pattern is probably repeated periodically. Thus we can expect a poor partition result when uniformly partitioning a one-day-span but a much better one for a period comprising several days, in particular when m matches the number of days. We will look into this issue further when we evaluate the partitioning strategies in chapter 10.