Predicate-size predicate partitioning

[EnricoMi]

The zero service provides predicate size statistics. A predicate partitioner could bin predicates by size, trying to achieve equal-size predicate partitions. This would be useful to partition long-tail schemata where we do not want a fixed number of predicates in partitions or all predicates in a singleton partition. The goal could be to have partitions with size of the largest predicate or a multiple of it.

The result partitioning can be combined with a uid partitioning to increase parallelism if needed.

A binning algorithm could desending-sort predicates by size and put the next predicate into the partition with the smallest amount of remaining space that can still accommodate it. If there is no such partition, create a new one. Hence maximum partition size must be greater or equal to the largest predicate. With a long-tail schema, smaller predicates will fill up the remaining space. They may be a small number of partitions with many tail-predicates, though.