Administration: Omnidex Indexing

Indexing Concepts

Aggregations

Omnidex indexing is valuable for processing aggregations. Aggregations involve the use of the COUNT, SUM, AVERAGE, MIN and MAX functions, and may or may not include a GROUP BY clause. Omnidex first considers the type of aggregation. Queries to request a COUNT(*) without a GROUP BY clause are the simplest queries. These are optimized differently than queries that use a GROUP BY clause or use other aggregation functions.

Simple COUNT(*) Queries

Omnidex always maintains a count when accessing its indexes. Whether processing criteria or performing a table join, a count is always available. Counts that simply request COUNT(*) are subsequently the simplest to optimize.

The following query requests a simple count from a table.

  select        count(*)
    from        HOUSEHOLDS
    where       STATE = 'CO' and
                CITY = 'Boulder' and
                ZIP = '80301';

Omnidex will search its indexes to process the criteria, and it will have the count ready when it is complete. There is no additional work needed to return the count, and no access to the database is required. Since these types of queries are very common, this makes Omnidex especially fast as a count engine.

General Aggregations

Queries that aggregate data often use GROUP BY clauses or other aggregation functions. These queries can be optimized with Omnidex indexing as well. The following query shows counts from the HOUSEHOLDS table aggregated by STATE, CITY and ZIP.

  select        STATE, CITY, ZIP, count(*)
    from        HOUSEHOLDS
    group by    STATE, CITY, ZIP

An index has the advantage of already being in sorted order. If the STATE, CITY and ZIP columns are all individually indexed with Omnidex, then these three indexes will be used in combination to perform the aggregation. The STATE index will be scanned, followed by the CITY index, followed by the ZIP index. The aggregations can be returned without accessing the database.

Some aggregations benefit from multi-column indexes, such as a single index with the STATE, CITY and ZIP. This is more efficient than using three separate indexes, and is especially important if any of the columns contain fairly unique values. It is also required for aggregations that use the SUM, AVERAGE, MIN and MAX functions.

These techniques can be used even after other indexes have been used to process criteria. After the criteria has been processed, the isolated record pointers can be used as a filter when the aggregation index is scanned. This allows for great flexibility, since the columns referenced in criteria do not need to be included in the aggregation index. In the following query, the same aggregations are restricted to people in the United States.

  select        STATE, CITY, ZIP, count(*)
    from        HOUSEHOLDS
    where       COUNTRY = 'US'
    group by    STATE, CITY, ZIP

Omnidex will first qualify the rows in the United States, and will then allow those record pointers to be used as a filter against the aggregation index.

When to Avoid Indexes

As was seen with criteria and table joins, sometimes an index is not the fastest approach. Using an index for aggregations usually requires scanning the index from the beginning to the end, even when there is criteria. If the criteria isolates a very small number of rows, it would be faster to simply retrieve the rows and aggregate them directly, than to scan millions of entries in an index.

As a general rule of thumb, it is valuable to use indexing for processing criteria when aggregating greater than 0.01% of the rows in a table. Omnidex automatically follows this rule of thumb, and you may find that some queries don't use indexes for this very reason. Omnidex also allows this threshold to be configured. A setting called the AGGREGATION_THRESHOLD is set to 0.01% by default, but can be changed if needed.

The following article explains basic strategies for optimizing aggregations.

Additional Resources

See also:

 
Back to top
admin/indexing/concepts/aggregations.txt ยท Last modified: 2016/06/28 22:38 (external edit)