Improve logic when to write column indexes

[asfimport]

Currently, we always write column indexes. In case of the data is ordered (ASCENDING or DESCENDING) the filtering would highly benefit from column indexes. While, if the data is UNORDERED it is not obvious if ordering based on column indexes would make sense. For example if the data is random then the min/max values of the different pages might be close to each other so in most cases filtering based on these values would not drop any of the pages. In the other hand UNORDERED values does not mean that the values are random. It can happen that the values are clustered or semi-ordered. We shall discover these cases somehow before writing the column indexes and write only if the min/max values for the pages do not overlap too much.

Another simple case if we have only one page. In this case writing column indexes is useless. 

Reporter: Gabor Szadovszky / @gszadovszky Assignee: Gabor Szadovszky / @gszadovszky

Related issues:

• Column indexes (depends upon)
• Benchmark filtering column-indexes (depends upon)

_{Note: This issue was originally created as PARQUET-1415. Please see the migration documentation for further details.}

[asfimport]

Gabor Szadovszky / @gszadovszky: Decide if we need to work on this issue based on the results of the benchmarks.

[asfimport]

Gabor Szadovszky / @gszadovszky: Decreased priority after benchmarks show that there is no significant overhead of using column-indexes in case of random data.