Spatial

Calcite is aiming to implement OpenGIS Simple Features Implementation Specification for SQL, version 1.2.1, a standard implemented by spatial databases such as PostGIS and H2GIS.

We also aim to add optimizer support for spatial indexes and other forms of query optimization.

Introduction

A spatial database is a database that is optimized for storing and query data that represents objects defined in a geometric space.

Calcite’s support for spatial data includes:

  • A GEOMETRY data type and sub-types including POINT, LINESTRING and POLYGON
  • Spatial functions (prefixed ST_; we have implemented about 35 of the 150 in the OpenGIS specification)

and will at some point also include query rewrites to use spatial indexes.

Enabling spatial support

Though the GEOMETRY data type is built-in, the functions are not enabled by default. You need to add fun=spatial to the JDBC connect string to enable the functions. For example, sqlline:

$ ./sqlline
> !connect jdbc:calcite:fun=spatial "sa" ""
SELECT ST_PointFromText('POINT(-71.064544 42.28787)');
+-------------------------------+
| EXPR$0                        |
+-------------------------------+
| {"x":-71.064544,"y":42.28787} |
+-------------------------------+
1 row selected (0.323 seconds)

Query rewrites

One class of rewrites uses Hilbert space-filling curves. Suppose that a table has columns x and y denoting the position of a point and also a column h denoting the distance of that point along a curve. Then a predicate involving distance of (x, y) from a fixed point can be translated into a predicate involving ranges of h.

Suppose we have a table with the locations of restaurants:

CREATE TABLE Restaurants (
  INT id NOT NULL PRIMARY KEY,
  VARCHAR(30) name,
  VARCHAR(20) cuisine,
  INT x NOT NULL,
  INT y NOT NULL,
  INT h  NOT NULL DERIVED (ST_Hilbert(x, y)))
SORT KEY (h);

The optimizer requires that h is the position on the Hilbert curve of point (x, y), and also requires that the table is sorted on h. The DERIVED and SORT KEY clauses in the DDL syntax are invented for the purposes of this example, but a clustered table with a CHECK constraint would work just as well.

The query

SELECT *
FROM Restaurants
WHERE ST_DWithin(ST_Point(x, y), ST_Point(10.0, 20.0), 6)

can be rewritten to

SELECT *
FROM Restaurants
WHERE (h BETWEEN 36496 AND 36520
    OR h BETWEEN 36456 AND 36464
    OR h BETWEEN 33252 AND 33254
    OR h BETWEEN 33236 AND 33244
    OR h BETWEEN 33164 AND 33176
    OR h BETWEEN 33092 AND 33100
    OR h BETWEEN 33055 AND 33080
    OR h BETWEEN 33050 AND 33053
    OR h BETWEEN 33033 AND 33035)
AND ST_DWithin(ST_Point(x, y), ST_Point(10.0, 20.0), 6)

The rewritten query contains a collection of ranges on h followed by the original ST_DWithin predicate. The range predicates are evaluated first and are very fast because the table is sorted on h.

Here is the full set of transformations:

Description Expression
Test whether a constant rectangle (X, X2, Y, Y2) contains a point (a, b)

Rewrite to use Hilbert index
ST_Contains(​ST_Rectangle(​X, X2, Y, Y2), ST_Point(a, b)))

h BETWEEN C1 AND C2
OR …
OR h BETWEEN C2k AND C2k+1
Test whether a constant geometry G contains a point (a, b)

Rewrite to use bounding box of constant geometry, which is also constant, then rewrite to Hilbert range(s) as above
ST_Contains(​ST_Envelope(​G), ST_Point(a, b))

ST_Contains(​ST_Rectangle(​X, X2, Y, Y2), ST_Point(a, b)))
Test whether a point (a, b) is within a buffer around a constant point (X, Y)

Special case of previous, because buffer is a constant geometry
ST_Contains(​ST_Buffer(​ST_Point(a, b), D), ST_Point(X, Y))
Test whether a point (a, b) is within a constant distance D of a constant point (X, Y)

First, convert to buffer, then use previous rewrite for constant geometry
ST_DWithin(​ST_Point(a, b), ST_Point(X, Y), D))

ST_Contains(​ST_Buffer(​ST_Point(​X, Y), D), ST_Point(a, b))
Test whether a constant point (X, Y) is within a constant distance D of a point (a, b)

Reverse arguments of call to ST_DWithin, then use previous rewrite
ST_DWithin(​ST_Point(X, Y), ST_Point(a, b), D))

ST_Contains(​ST_Buffer(​ST_Point(​X, Y), D), ST_Point(a, b))

In the above, a and b are variables, X, X2, Y, Y2, D and G are constants.

Many rewrites are inexact: there are some points where the predicate would return false but the rewritten predicate returns true. For example, a rewrite might convert a test whether a point is in a circle to a test for whether the point is in the circle’s bounding square. These rewrites are worth performing because they are much quicker to apply, and often allow range scans on the Hilbert index. But for safety, Calcite applies the original predicate, to remove false positives.

Acknowledgements

Calcite’s OpenGIS implementation uses the Esri geometry API. Thanks for the help we received from their community.

While developing this feature, we made extensive use of the PostGIS documentation and tests, and the H2GIS documentation, and consulted both as reference implementations when the specification wasn’t clear. Thank you to these awesome projects.