[Feature Request]: Secondary index

[yingfeng]

Secondary index is used for numeric filtering. It is composed of two parts:

1. The data of each numeric column is stored in an inverted sorted form, with a compressed format.
2. Another in-memory part of index data which is based on pgm, it could provide very fast approximate range query with bounded error, which has already been added into the repository.

The mechanism of range filtering of secondary index is as follows:

1. Query the pgm index to get the bounded range.
2. Scan the raw index data according to bounded range to get the RowIDs of the query filter.

[yangzq50]

My proposal for the implementation:

1. When building secondary index for a numeric column in the table, we can build one secondary index structure for each segment of the table.

2. Every secondary index structure should contain 3 parts:
   2.1. pgm: a data structure which provides approximate range
   2.2. vector_value: sorted array of all the values of the chosen column in the segment (necessary for getting exact range)
   2.3. vector_offset: array of segment offsets for each value in vector_value (necessary for getting filtering results)

3. vector_value and vector_offset can be split into parts with max size DEFAULT_BLOCK_CAPACITY

4. Steps for range search of one column with secondary index:
   4.1. Get an approximate range from pgm
   4.2. Get the exact range by binary search in vector_value
   4.3. Generate a bitmask (when the size of the exact range is too big) or a vector of chosen offsets

[yingfeng]

4.2 is not necessary because pgm has already told the approximate range, as a result 4.2 is acturally a linear scan

4. Steps for range search of one column with secondary index:
   4.1. Get an approximate range from pgm
   4.2. Get the exact range by binary search in vector_value
   4.3. Generate a bitmask (when the size of the exact range is too big) or a vector of chosen offsets

[yangzq50]

subtask #476 and subtask #477 are deferred to #637