Current PostgreSQL versions (including 9.6) suffer from a query optimizer weakness regarding TOAST tables. TOAST tables are a kind of "extension room" used to store large (in the sense of data size) values that do not fit into normal data pages (like long texts, images or complex geometries with lots of vertices), see the PostgreSQL Documentation for TOAST for more information).
The problem appears if you happen to have a table with rather large geometries, but not too many rows of them (like a table containing the boundaries of all European countries in high resolution). Then the table itself is small, but it uses lots of TOAST space. In our example case, the table itself had about 80 rows and used only 3 data pages, but the TOAST table used 8225 pages.
Now issue a query where you use the geometry operator && to search for a bounding box that matches only very few of those rows. Now the query optimizer sees that the table has only 3 pages and 80 rows. It estimates that a sequential scan on such a small table is much faster than using an index. And so it decides to ignore the GIST index. Usually, this estimation is correct. But in our case, the && operator has to fetch every geometry from disk to compare the bounding boxes, thus reading all TOAST pages, too.
To see whether your suffer from this issue, use the "EXPLAIN ANALYZE" postgresql command. For more information and the technical details, you can read the thread on the PostgreSQL performance mailing list: http://archives.postgresql.org/pgsql-performance/2005-02/msg00030.php
and newer thread on PostGIS https://lists.osgeo.org/pipermail/postgis-devel/2017-June/026209.html
The PostgreSQL people are trying to solve this issue by making the query estimation TOAST-aware. For now, here are two workarounds:
The first workaround is to force the query planner to use the index. Send "SET enable_seqscan TO off;" to the server before issuing the query. This basically forces the query planner to avoid sequential scans whenever possible. So it uses the GIST index as usual. But this flag has to be set on every connection, and it causes the query planner to make misestimations in other cases, so you should "SET enable_seqscan TO on;" after the query.
The second workaround is to make the sequential scan as fast as the query planner thinks. This can be achieved by creating an additional column that "caches" the bbox, and matching against this. In our example, the commands are like:
SELECT AddGeometryColumn('myschema','mytable','bbox','4326','GEOMETRY','2'); UPDATE mytable SET bbox = ST_Envelope(ST_Force2D(geom));
Now change your query to use the && operator against bbox instead of geom_column, like:
SELECT geom_column FROM mytable WHERE bbox && ST_SetSRID('BOX3D(0 0,1 1)'::box3d,4326);
Of course, if you change or add rows to mytable, you have to keep the bbox "in sync". The most transparent way to do this would be triggers, but you also can modify your application to keep the bbox column current or run the UPDATE query above after every modification.
For tables that are mostly read-only, and where a single index is used for the majority of queries, PostgreSQL offers the CLUSTER command. This command physically reorders all the data rows in the same order as the index criteria, yielding two performance advantages: First, for index range scans, the number of seeks on the data table is drastically reduced. Second, if your working set concentrates to some small intervals on the indices, you have a more efficient caching because the data rows are spread along fewer data pages. (Feel invited to read the CLUSTER command documentation from the PostgreSQL manual at this point.)
However, currently PostgreSQL does not allow clustering on PostGIS GIST indices because GIST indices simply ignores NULL values, you get an error message like:
lwgeom=# CLUSTER my_geom_index ON my_table; ERROR: cannot cluster when index access method does not handle null values HINT: You may be able to work around this by marking column "geom" NOT NULL.
As the HINT message tells you, one can work around this deficiency by adding a "not null" constraint to the table:
lwgeom=# ALTER TABLE my_table ALTER COLUMN geom SET not null; ALTER TABLE
Of course, this will not work if you in fact need NULL values in your geometry column. Additionally, you must use the above method to add the constraint, using a CHECK constraint like "ALTER TABLE blubb ADD CHECK (geometry is not null);" will not work.
Sometimes, you happen to have 3D or 4D data in your table, but always access it using OpenGIS compliant ST_AsText() or ST_AsBinary() functions that only output 2D geometries. They do this by internally calling the ST_Force2D() function, which introduces a significant overhead for large geometries. To avoid this overhead, it may be feasible to pre-drop those additional dimensions once and forever:
UPDATE mytable SET geom = ST_Force2D(geom); VACUUM FULL ANALYZE mytable;
Note that if you added your geometry column using AddGeometryColumn() there'll be a constraint on geometry dimension. To bypass it you will need to drop the constraint. Remember to update the entry in the geometry_columns table and recreate the constraint afterwards.
In case of large tables, it may be wise to divide this UPDATE into smaller portions by constraining the UPDATE to a part of the table via a WHERE clause and your primary key or another feasible criteria, and running a simple "VACUUM;" between your UPDATEs. This drastically reduces the need for temporary disk space. Additionally, if you have mixed dimension geometries, restricting the UPDATE by "WHERE dimension(geom)>2" skips re-writing of geometries that already are in 2D.