[postgis-devel] Optimizing distance routines in PostGIS

[Sandro Santilli]

On Mon, Oct 31, 2016 at 06:17:34PM -0400, Daniel Baston wrote:
> I recently ported the IndexedFacetDistance class from JTS to GEOS.
> This class provides optimized distance routines that work by
> constructing Rtree indexes over both of the input geometries, and
> performing distance calculations only between the nearest portions
> (facets) of the two inputs.  This can produce massive performance
> improvements on distance calculations.  (An example discussed in trac
> #3587 goes from about 42 seconds to 0.1 seconds.)  Additionally, the
> indexes can be retained (much like a PreparedGeometry), for the case
> where many inputs need to be tested against a single large geometry
> (this last bit hasn't been exposed in the C API).
> 
> Is this something we should explore using in ST_Distance / ST_DWithin,
> or maybe elsewhere in PostGIS?  It would have to be limited to types
> supported by GEOS (2D only, no curves), and we'd probably need to find
> some heuristics about when it's worth the effort (rather than a
> standard lwgeom_mindistance2d).

I've been working with distances recently, for a new snap algorithm
in librttopo, and found myself in need of not just a distance value
but also which element was closest (to a point, in my case).
The liblwgeom functions (in measures.c) seem to be returning the
detail in terms of "closest points", but that's still not enough
detail for what I needed.

Could the yet-to-be-exposed part of the API give additional details ?
To which level ?

--strk;