/pgfaceting

Faceted query acceleration for PostgreSQL using roaring bitmaps

Primary LanguagePLpgSQLBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

pgfaceting

PostgreSQL extension to quickly calculate facet counts using inverted index built with roaring bitmaps. Requires pg_roaringbitmap to be installed.

Faceting means counting number occurrences of each value in a result set for a set of attributes. Typical example of faceting is be a web shop where you can see how many items are remaining after filtering your search by red, green or blue, and how many when filtering by size small, medium or large.

Work on this project has been sponsored by Xenit.

Build and install

make install
make installcheck
psql -c "create extension roaringbitmap" -c "create extension pgfaceting"

Usage

pgfaceting creates and maintains two extra tables for your main table. tbl_facets contains for each facet and value combination a list of id values for rows containing that combination. The list is stored as a roaring bitmap for quick intersection and cardinality operations. Because updating this list is a heavy operation any changes to the main table get stored in tbl_facets_deltas as a combination of facet, value, id and +1 or -1 depending on the kind of update. A periodic maintenance job is responsible for merging deltas into the main facets table.

Currently only 32bit integer id columns are supported. When pg_roaringbitmap adds support for 64bit bitmaps then int8 and possibly ctid could be supported.

Adding faceting to a table

SELECT faceting.add_faceting_to_table(
    'documents',
    key => 'id',
    facets => array[
        faceting.datetrunc_facet('created', 'month'),
        faceting.datetrunc_facet('finished', 'month'),
        faceting.plain_facet('category_id'),
        faceting.plain_facet('type'),
        faceting.bucket_facet('size', buckets => array[0,1000,5000,10000,50000,100000,500000])
    ]
);

The add_faceting_to_table() adds the facets tables and populates the contents. It takes an array of facets to extract from each row.

  • plain_facet(col name) - Takes the column value as is as the facet value.
  • datetrunc_facet(col name, precision text) - Applies a date_trunc function on a column to get the facet value. Useful for timebucketing (yearly, monthly, etc.)
  • bucket_facet(col name, buckets anyarray) - Assigns a continuous variable (price, weight, etc.) to a set of buckets and stores the index of the chosen bucket as the facet value.

For merging changes create a periodic job that runs:

CALL faceting.run_maintenance();

This will run delta merging on all faceted tables. There is also a function for maintaining a single table:

SELECT faceting.merge_deltas('documents'::regclass);

Querying facets

Getting top 10 values for each kind of facet:

SELECT * FROM faceting.top_values('documents'::regclass, n => 10);

We can also filter by some facets and get the results of other facets:

SELECT * FROM faceting.count_results('documents'::regclass,
                                     filters => array[row('category_id', '24'),
                                                      row('type', 'image/jpeg')]::faceting.facet_filter[]);

For advanced usage the inverted index tables can be accessed directly.

WITH lookup AS (
    SELECT id >> 20 AS chunk_id, rb_build_agg(id) postinglist
    FROM documents
    WHERE ...
)
SELECT facet_id, facet_value, sum(rb_and_cardinality(flt.postinglist, fct.postinglist))
FROM lookup flt JOIN documents_facets USING (chunk_id)
GROUP BY 1, 2;

How fast is it

Calculating facets for 61% of rows in 100M row table:

-- 24 vcore parallel seq scan
postgres=# SELECT facet_name, count(distinct facet_value), sum(cardinality)
FROM (SELECT facet_name, facet_value, COUNT(*) cardinality
      FROM test2.documents d, LATERAL (VALUES
            ('created', date_trunc('month', created)::text),
            ('finished', date_trunc('month', finished)::text),
            ('type', type::text),
            ('size', width_bucket(size, array[0,1000,5000,10000,50000,100000,500000])::text)
          ) t(facet_name, facet_value)
      WHERE category_id = 24
      GROUP BY 1, 2) count_results
GROUP BY 1;
 facet_name | count |   sum    
------------+-------+----------
 created    |   154 | 60812252
 finished   |   154 | 60812252
 size       |     7 | 60812252
 type       |     8 | 60812252
(4 rows)

Time: 18440.061 ms (00:18.440)

-- Single core only
postgres=# SET max_parallel_workers_per_gather = 0;
SET
Time: 0.206 ms
postgres=# SELECT facet_name, count(distinct facet_value), sum(cardinality)
FROM (SELECT facet_name, facet_value, COUNT(*) cardinality
      FROM test2.documents d, LATERAL (VALUES
            ('created', date_trunc('month', created)::text),
            ('finished', date_trunc('month', finished)::text),
            ('type', type::text),
            ('size', width_bucket(size, array[0,1000,5000,10000,50000,100000,500000])::text)
          ) t(facet_name, facet_value)
      WHERE category_id = 24
      GROUP BY 1, 2) count_results
GROUP BY 1;
 facet_name | count |   sum    
------------+-------+----------
 created    |   154 | 60812252
 finished   |   154 | 60812252
 size       |     7 | 60812252
 type       |     8 | 60812252
(4 rows)

Time: 222019.758 ms (03:42.020)

-- Using facets index
postgres=# SELECT facet_name, count(distinct facet_value), sum(cardinality)
FROM faceting.count_results('documents'::regclass,
    filters => array[row('category_id', 24)]::faceting.facet_filter[])
GROUP BY 1;
 facet_name | count |   sum    
------------+-------+----------
 created    |   154 | 60812252
 finished   |   154 | 60812252
 size       |     7 | 60812252
 type       |     8 | 60812252
(4 rows)

 Time: 155.228 ms