TrieRangeQuery): Schindler, U, Diepenbroek, M, 2008. Generic XML-based Framework for Metadata Portals. Computers & Geosciences 34 (12), 1947-1955. doi:10.1016/j.cageo.2008.02.023
A quote from this paper: Because Apache Lucene is a full-text search engine and not a conventional database, it cannot handle numerical ranges (e.g., field value is inside user defined bounds, even dates are numerical values). We have developed an extension to Apache Lucene that stores the numerical values in a special string-encoded format with variable precision (all numerical values like doubles, longs, floats, and ints are converted to lexicographic sortable string representations and stored with different precisions (for a more detailed description of how the values are stored, see {@link NumericUtils}). A range is then divided recursively into multiple intervals for searching: The center of the range is searched only with the lowest possible precision in the trie, while the boundaries are matched more exactly. This reduces the number of terms dramatically.
For the variant that stores long values in 8 different precisions (each reduced by 8 bits) that uses a lowest precision of 1 byte, the index contains only a maximum of 256 distinct values in the lowest precision. Overall, a range could consist of a theoretical maximum of 7*255*2 + 255 = 3825 distinct terms (when there is a term for every distinct value of an 8-byte-number in the index and the range covers almost all of them; a maximum of 255 distinct values is used because it would always be possible to reduce the full 256 values to one term with degraded precision). In practice, we have seen up to 300 terms in most cases (index with 500,000 metadata records and a uniform value distribution).
You can choose any precisionStep when encoding values. Lower step values mean more precisions and so more terms in index (and index gets larger). On the other hand, the maximum number of terms to match reduces, which optimized query speed. The formula to calculate the maximum term count is:
n = [ (bitsPerValue/precisionStep - 1) * (2^precisionStep - 1 ) * 2 ] + (2^precisionStep - 1 )
(this formula is only correct, when bitsPerValue/precisionStep is an integer; in other cases, the value must be rounded up and the last summand must contain the modulo of the division as precision step). For longs stored using a precision step of 4, n = 15*15*2 + 15 = 465, and for a precision step of 2, n = 31*3*2 + 3 = 189. But the faster search speed is reduced by more seeking in the term enum of the index. Because of this, the ideal precisionStep value can only be found out by testing. Important: You can index with a lower precision step value and test search speed using a multiple of the original step value.
Good values for precisionStep are depending on usage and data type:
precisionStep is given. precisionStep). Using {@link NumericField NumericFields} for sortingis ideal, because building the field cache is much faster than with text-only numbers. These fields have one term per value and therefore also work with term enumeration for building distinct lists (e.g. facets / preselected values to search for). Sorting is also possible with range query optimized fields using one of the above precisionSteps. Comparisons of the different types of RangeQueries on an index with about 500,000 docs showed that {@link TermRangeQuery} in boolean rewrite mode (with raised {@link BooleanQuery} clause count)took about 30-40 secs to complete, {@link TermRangeQuery} in constant score filter rewrite mode took 5 secsand executing this class took <100ms to complete (on an Opteron64 machine, Java 1.5, 8 bit precision step). This query type was developed for a geographic portal, where the performance for e.g. bounding boxes or exact date/time stamps is important.
@since 2.9 | |