This infographic visually compares the difference between running a PostGIS comparison query like ST_Intersects on a large shape versus a subdivided version of that large shape. Click to embiggen.
Hundreds of GIS intersection comparisons are completed every hour on Chicago Cityscape.*
People are looking at, say, a map of the South Shore community area. That “Place” page then grabs all of the building permits, building violations, business licenses, and other “feature layers” that are stored as points.
A classic “point in polygon” comparison is made using the
ST_Intersects(place_geometry, permits_geometry) function.
This has worked very well for several years.
But as Chicago Cityscape handles larger shapes – they come from users drawing their own, large shapes, and from large shapes like the downtown Chicago area – this query doesn’t cut it.
Setting indexes on the geometry is imperative, but it’s not the end of the to optimize performance. That’s because the index of the geometry is a rectangular bounding box (which is also called an “envelope” in GIS) that contains the entire shape of the South Shore community area.
The downtown Chicago area, however, is not even the largest shape I have. That belongs to the new Place, “Neighborhood Opportunity Fund investment zones” (NOF). Combined, they cover 75 square miles of Chicago. Downtown is only 7.7 square miles.
After I added the NOF map and tested its Place page, it “crashed” my server, metaphorically speaking. The query to just count the number of building permits in the area would take about five minutes.
There had to be a better way; in the meantime, however, I divided the NOF map into the West and South sections. This hardly improved the counting time.
Thankfully, today, I saw a tweet from Paul Ramsey linking to his blog that linked to his slides from a recent presentation about the use of PostgreSQL to store and manipulate GIS data.
In it he explained how the
ST_Subdividefunction worked. I’m going to demonstrate it using graphics from my own maps.
A normal intersection comparison, using
ST_Intersects(place_geometry, permits_geometry) in a query creates a bounding box (envelope) around each geometry and quickly determines whether the two envelopes overlap. If they do, then it checks again to see if the actual geometries overlap. If they do, that data is returned as a response to your query.
When your two datasets are massive, like the NOF zones, which collectively cover 1/3rd of Chicago, and the building permits, which are found across the entire city…well, that led to the five minutes counting time.
ST_Subdivide. To use it properly you would run it against your existing geometry and store the much smaller shapes, derived from the big shape, in a new table. I applied the function to all the 22,203 maps that Chicago Cityscape has and stored their unique IDs and subdivided geometries in a new “lookup” table.
Now, any time I want to compare the building permits against the NOF, the building permits are instead compared to the small shapes that were subdivided.
Chicago Cityscape uses a single table (created as a materialized view) to combine all 22,203 maps. Each map is stored in a source table (for example, there’s a table to hold the 77 community areas) and the materialized view runs once a day to combine all of the maps in the source tables. This ensures our data is managed well: different source tables can hold different information, and the single table holds only the name, type, and geometry of the source tables, for faster comparison. Each entry in the single table also has a “slug”, its unique identifier.
Thus, the materialized view of the subdivided maps is created from the aforementioned single table, using this query:
create materialized view view_places_subdivided as
select gid || '_' || random() as gid, slug, st_subdivide(geom) as geom
The “gid” is designed to create a new unique ID field, as the slug field will be repeated for every subdivided of each map. A unique ID field is necessary if you want to refresh the materialized view concurrently (to allow for other queries to access the materialized view while it’s being refreshed).
* The results are cached for a few hours, because the feature layers change 1-2 times per day and at different times each day, so the limited duration cache accommodates that. Ideally I would code a way to invalidate the cache when the feature layer data is updated.