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 manyrows 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 postgres 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(the_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 "the_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 the_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 the_geom = ST_Force2D(the_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(the_geom)>2" skips re-writing of geometries that already are in 2D.
Tuning for PostGIS is much like tuning for any PostgreSQL workload. The only additional note to keep in mind is that geometries and rasters are heavy so memory related optimizations generally have more of an impact on PostGIS than other types of PostgreSQL queries.
For general details about optimizing PostgreSQL, refer to Tuning your PostgreSQL Server.
For PostgreSQL 9.4+ all these can be set at the server level without touching postgresql.conf or postgresql.auto.conf
by using the
ALTER SYSTEM.. command.
ALTER SYSTEM SET work_mem = '256MB'; -- this will force, non-startup configs to take effect for new connections SELECT pg_reload_conf(); -- show current setting value -- use SHOW ALL to see all settings SHOW work_mem;
In addition to these settings, PostGIS also has some custom settings which you can find listed in Section 8.2, “PostGIS Grand Unified Custom Variables (GUCs)”.
These settings are configured in postgresql.conf:
This is generally used for table partitioning. The default for this is set to "partition" which is ideal for PostgreSQL 8.4 and above since it will force the planner to only analyze tables for constraint consideration if they are in an inherited hierarchy and not pay the planner penalty otherwise.
Default: ~128MB in PostgreSQL 9.6
Set to about 25% to 40% of available RAM. On windows you may not be able to set as high.
max_worker_processes This setting is only available for PostgreSQL 9.4+. For PostgreSQL 9.6+ this setting has additional importance in that it controls the max number of processes you can have for parallel queries.
Sets the maximum number of background processes that the system can support. This parameter can only be set at server start.
work_mem (the memory used for sort operations and complex queries)
Adjust up for large dbs, complex queries, lots of RAM
Adjust down for many concurrent users or low RAM.
If you have lots of RAM and few developers:
SET work_mem TO '256MB';
maintenance_work_mem (used for VACUUM, CREATE INDEX, etc.)
Generally too low - ties up I/O, locks objects while swapping memory
Recommend 32MB to 1GB on production servers w/lots of RAM, but depends on the # of concurrent users. If you have lots of RAM and few developers:
SET maintenance_work_mem TO '1GB';
This setting is only available for PostgreSQL 9.6+ and will only affect PostGIS 2.3+, since only PostGIS 2.3+ supports parallel queries.
If set to higher than 0, then some queries such as those involving relation functions like
ST_Intersects can use multiple processes and can run
more than twice as fast when doing so. If you have a lot of processors to spare, you should change the value of this to as many processors as you have.
Also make sure to bump up
max_worker_processes to at least as high as this number.
Sets the maximum number of workers that can be started
by a single
Parallel workers are taken from the pool of processes
Note that the requested number of workers may not
actually be available at run time. If this occurs, the
plan will run with fewer workers than expected, which may
be inefficient. Setting this value to 0, which is the
default, disables parallel query execution.