Skip to content
This documentation is sourced from a third-party project and is not maintained by pgEdge.

Spatial Indexes

Spatial Indexes

Spatial indexes make using a spatial database for large data sets possible. Without indexing, a search for features requires a sequential scan of every record in the database. Indexing speeds up searching by organizing the data into a structure which can be quickly traversed to find matching records.

The B-tree index method commonly used for attribute data is not very useful for spatial data, since it only supports storing and querying data in a single dimension. Data such as geometry (which has 2 or more dimensions) requires an index method that supports range query across all the data dimensions. One of the key advantages of PostgreSQL for spatial data handling is that it offers several kinds of index methods which work well for multi-dimensional data: GiST, BRIN and SP-GiST indexes.

  • GiST (Generalized Search Tree) indexes break up data into "things to one side", "things which overlap", "things which are inside" and can be used on a wide range of data-types, including GIS data. PostGIS uses an R-Tree index implemented on top of GiST to index spatial data. GiST is the most commonly-used and versatile spatial index method, and offers very good query performance.
  • BRIN (Block Range Index) indexes operate by summarizing the spatial extent of ranges of table records. Search is done via a scan of the ranges. BRIN is only appropriate for use for some kinds of data (spatially sorted, with infrequent or no update). But it provides much faster index create time, and much smaller index size.
  • SP-GiST (Space-Partitioned Generalized Search Tree) is a generic index method that supports partitioned search trees such as quad-trees, k-d trees, and radix trees (tries).

Spatial indexes store only the bounding box of geometries. Spatial queries use the index as a primary filter to quickly determine a set of geometries potentially matching the query condition. Most spatial queries require a secondary filter that uses a spatial predicate function to test a more specific spatial condition. For more information on queying with spatial predicates see Using Spatial Indexes.

See also the PostGIS Workshop section on spatial indexes, and the PostgreSQL manual.

GiST Indexes

GiST stands for "Generalized Search Tree" and is a generic form of indexing for multi-dimensional data. PostGIS uses an R-Tree index implemented on top of GiST to index spatial data. GiST is the most commonly-used and versatile spatial index method, and offers very good query performance. Other implementations of GiST are used to speed up searches on all kinds of irregular data structures (integer arrays, spectral data, etc) which are not amenable to normal B-Tree indexing. For more information see the PostgreSQL manual.

Once a spatial data table exceeds a few thousand rows, you will want to build an index to speed up spatial searches of the data (unless all your searches are based on attributes, in which case you'll want to build a normal index on the attribute fields).

The syntax for building a GiST index on a "geometry" column is as follows:

CREATE INDEX [indexname] ON [tablename] USING GIST ( [geometryfield] );

The above syntax will always build a 2D-index. To get the an n-dimensional index for the geometry type, you can create one using this syntax:

CREATE INDEX [indexname] ON [tablename] USING GIST ([geometryfield] gist_geometry_ops_nd);

Building a spatial index is a computationally intensive exercise. It also blocks write access to your table for the time it creates, so on a production system you may want to do in in a slower CONCURRENTLY-aware way:

CREATE INDEX CONCURRENTLY [indexname] ON [tablename] USING GIST ( [geometryfield] );

After building an index, it is sometimes helpful to force PostgreSQL to collect table statistics, which are used to optimize query plans:

VACUUM ANALYZE [table_name] [(column_name)];

BRIN Indexes

BRIN stands for "Block Range Index". It is a general-purpose index method introduced in PostgreSQL 9.5. BRIN is a lossy index method, meaning that a secondary check is required to confirm that a record matches a given search condition (which is the case for all provided spatial indexes). It provides much faster index creation and much smaller index size, with reasonable read performance. Its primary purpose is to support indexing very large tables on columns which have a correlation with their physical location within the table. In addition to spatial indexing, BRIN can speed up searches on various kinds of attribute data structures (integer, arrays etc). For more information see the PostgreSQL manual.

Once a spatial table exceeds a few thousand rows, you will want to build an index to speed up spatial searches of the data. GiST indexes are very performant as long as their size doesn't exceed the amount of RAM available for the database, and as long as you can afford the index storage size, and the cost of index update on write. Otherwise, for very large tables BRIN index can be considered as an alternative.

A BRIN index stores the bounding box enclosing all the geometries contained in the rows in a contiguous set of table blocks, called a block range. When executing a query using the index the block ranges are scanned to find the ones that intersect the query extent. This is efficient only if the data is physically ordered so that the bounding boxes for block ranges have minimal overlap (and ideally are mutually exclusive). The resulting index is very small in size, but is typically less performant for read than a GiST index over the same data.

Building a BRIN index is much less CPU-intensive than building a GiST index. It's common to find that a BRIN index is ten times faster to build than a GiST index over the same data. And because a BRIN index stores only one bounding box for each range of table blocks, it's common to use up to a thousand times less disk space than a GiST index.

You can choose the number of blocks to summarize in a range. If you decrease this number, the index will be bigger but will probably provide better performance.

For BRIN to be effective, the table data should be stored in a physical order which minimizes the amount of block extent overlap. It may be that the data is already sorted appropriately (for instance, if it is loaded from another dataset that is already sorted in spatial order). Otherwise, this can be accomplished by sorting the data by a one-dimensional spatial key. One way to do this is to create a new table sorted by the geometry values (which in recent PostGIS versions uses an efficient Hilbert curve ordering):

CREATE TABLE table_sorted AS
   SELECT * FROM table  ORDER BY geom;

Alternatively, data can be sorted in-place by using a GeoHash as a (temporary) index, and clustering on that index:

CREATE INDEX idx_temp_geohash ON table
    USING btree (ST_GeoHash( ST_Transform( geom, 4326 ), 20));
CLUSTER table USING idx_temp_geohash;

The syntax for building a BRIN index on a geometry column is:

CREATE INDEX [indexname] ON [tablename] USING BRIN ( [geome_col] );

The above syntax builds a 2D index. To build a 3D-dimensional index, use this syntax:

CREATE INDEX [indexname] ON [tablename]
    USING BRIN ([geome_col] brin_geometry_inclusion_ops_3d);

You can also get a 4D-dimensional index using the 4D operator class:

CREATE INDEX [indexname] ON [tablename]
    USING BRIN ([geome_col] brin_geometry_inclusion_ops_4d);

The above commands use the default number of blocks in a range, which is 128. To specify the number of blocks to summarise in a range, use this syntax

CREATE INDEX [indexname] ON [tablename]
    USING BRIN ( [geome_col] ) WITH (pages_per_range = [number]);

Keep in mind that a BRIN index only stores one index entry for a large number of rows. If your table stores geometries with a mixed number of dimensions, it's likely that the resulting index will have poor performance. You can avoid this performance penalty by choosing the operator class with the least number of dimensions of the stored geometries

The geography datatype is supported for BRIN indexing. The syntax for building a BRIN index on a geography column is:

CREATE INDEX [indexname] ON [tablename] USING BRIN ( [geog_col] );

The above syntax builds a 2D-index for geospatial objects on the spheroid.

Currently, only "inclusion support" is provided, meaning that just the &&, ~ and @ operators can be used for the 2D cases (for both geometry and geography), and just the &&& operator for 3D geometries. There is currently no support for kNN searches.

An important difference between BRIN and other index types is that the database does not maintain the index dynamically. Changes to spatial data in the table are simply appended to the end of the index. This will cause index search performance to degrade over time. The index can be updated by performing a VACUUM, or by using a special function brin_summarize_new_values(regclass). For this reason BRIN may be most appropriate for use with data that is read-only, or only rarely changing. For more information refer to the manual.

To summarize using BRIN for spatial data:

  • Index build time is very fast, and index size is very small.
  • Index query time is slower than GiST, but can still be very acceptable.
  • Requires table data to be sorted in a spatial ordering.
  • Requires manual index maintenance.
  • Most appropriate for very large tables, with low or no overlap (e.g. points), which are static or change infrequently.
  • More effective for queries which return relatively large numbers of data records.

SP-GiST Indexes

SP-GiST stands for "Space-Partitioned Generalized Search Tree" and is a generic form of indexing for multi-dimensional data types that supports partitioned search trees, such as quad-trees, k-d trees, and radix trees (tries). The common feature of these data structures is that they repeatedly divide the search space into partitions that need not be of equal size. In addition to spatial indexing, SP-GiST is used to speed up searches on many kinds of data, such as phone routing, ip routing, substring search, etc. For more information see the PostgreSQL manual.

As it is the case for GiST indexes, SP-GiST indexes are lossy, in the sense that they store the bounding box enclosing spatial objects. SP-GiST indexes can be considered as an alternative to GiST indexes.

Once a GIS data table exceeds a few thousand rows, an SP-GiST index may be used to speed up spatial searches of the data. The syntax for building an SP-GiST index on a "geometry" column is as follows:

CREATE INDEX [indexname] ON [tablename] USING SPGIST ( [geometryfield] );

The above syntax will build a 2-dimensional index. A 3-dimensional index for the geometry type can be created using the 3D operator class:

CREATE INDEX [indexname] ON [tablename] USING SPGIST ([geometryfield] spgist_geometry_ops_3d);

Building a spatial index is a computationally intensive operation. It also blocks write access to your table for the time it creates, so on a production system you may want to do in in a slower CONCURRENTLY-aware way:

CREATE INDEX CONCURRENTLY [indexname] ON [tablename] USING SPGIST ( [geometryfield] );

After building an index, it is sometimes helpful to force PostgreSQL to collect table statistics, which are used to optimize query plans:

VACUUM ANALYZE [table_name] [(column_name)];

An SP-GiST index can accelerate queries involving the following operators:

  • <<, &<, &>, >>, <<|, &<|, |&>, |>>, &&, @>, <@, and ~=, for 2-dimensional indexes,
  • &/&, ~==, @>>, and <<@, for 3-dimensional indexes.

There is no support for kNN searches at the moment.

Tuning Index Usage

Ordinarily, indexes invisibly speed up data access: once an index is built, the PostgreSQL query planner automatically decides when to use it to improve query performance. But there are some situations where the planner does not choose to use existing indexes, so queries end up using slow sequential scans instead of a spatial index.

If you find your spatial indexes are not being used, there are a few things you can do:

  • Examine the query plan and check your query actually computes the thing you need. An erroneous JOIN, either forgotten or to the wrong table, can unexpectedly retrieve table records multiple times. To get the query plan, execute with EXPLAIN in front of the query.
  • Make sure statistics are gathered about the number and distributions of values in a table, to provide the query planner with better information to make decisions around index usage. VACUUM ANALYZE will compute both.

You should regularly vacuum your databases anyways. Many PostgreSQL DBAs run VACUUM as an off-peak cron job on a regular basis. - If vacuuming does not help, you can temporarily force the planner to use the index information by using the command SET ENABLE_SEQSCAN TO OFF;. This way you can check whether the planner is at all able to generate an index-accelerated query plan for your query. You should only use this command for debugging; generally speaking, the planner knows better than you do about when to use indexes. Once you have run your query, do not forget to run SET ENABLE_SEQSCAN TO ON; so that the planner will operate normally for other queries. - If SET ENABLE_SEQSCAN TO OFF; helps your query to run faster, your Postgres is likely not tuned for your hardware. If you find the planner wrong about the cost of sequential versus index scans try reducing the value of RANDOM_PAGE_COST in postgresql.conf, or use SET RANDOM_PAGE_COST TO 1.1;. The default value for RANDOM_PAGE_COST is 4.0. Try setting it to 1.1 (for SSD) or 2.0 (for fast magnetic disks). Decreasing the value makes the planner more likely to use index scans. - If SET ENABLE_SEQSCAN TO OFF; does not help your query, the query may be using a SQL construct that the Postgres planner is not yet able to optimize. It may be possible to rewrite the query in a way that the planner is able to handle. For example, a subquery with an inline SELECT may not produce an efficient plan, but could possibly be rewritten using a LATERAL JOIN.

For more information see the Postgres manual section on Query Planning.