Precision And Tolerance Internals

PostGIS does not have one global tolerance model for all geometry operations. Precision and tolerance are handled at the API boundary where they affect a specific algorithm.

Current GEOS-backed overlay functions expose precision through gridSize, not through a general tolerance argument. ST_Intersection, ST_Difference, ST_SymDifference, ST_UnaryUnion, ST_Union, and ST_Subdivide have gridSize variants in postgis/postgis.sql.in.

The manual documents the contract through overlay_gridsize_arg: when gridSize is supplied, GEOS 3.9 or later is required and result vertices are guaranteed to fall on the specified snap-rounded grid. For repeatable results, input vertices should already lie on the same grid. Use ST_ReducePrecision when inputs must first be made grid-aligned; use ST_SnapToGrid for simpler coordinate-grid snapping where validity preservation is not the key contract. At the C boundary, ST_ReducePrecision reaches GEOS through lwgeom_reduceprecision and GEOSGeom_setPrecision.

This is the modern replacement for the old ToleranceDiscussion idea of passing an implicit GEOS PrecisionModel through PostGIS. The precision model is now explicit per operation where overlay needs it.

ST_Snap(input, reference, tolerance) uses a distance tolerance to snap input segments and vertices to vertices in a reference geometry. It is a robustness tool for nearly coincident overlay inputs, but it uses heuristics and can skip candidate snaps. Snapping can also make an output invalid or non-simple, so it is not a general precision model.

Distance predicates such as ST_DWithin, ST_3DDWithin, and ST_DFullyWithin have explicit distance arguments. They are user-visible spatial predicates, not replacements for exact-topology predicates such as ST_Intersects. Geography predicates also have their own documented implementation tolerances for some relationship tests.

Other tolerance parameters are algorithm-specific:

• ST_RemoveRepeatedPoints treats vertices within the tolerance as duplicate consecutive points.
• ST_DelaunayTriangles, ST_VoronoiPolygons, and ST_VoronoiLines use tolerance to snap input points for triangulation or Voronoi construction.
• ST_CoverageInvalidEdges and ST_CoverageSimplify use tolerances for coverage validation and simplification.
• Clustering functions such as ST_ClusterDBSCAN and ST_ClusterWithin use a distance threshold for grouping, not for coordinate precision.

The topology extension has its own precision and tolerance surface. A topology created with CreateTopology(toponame, srid, prec) records precision in the topology metadata. Topology loaders and converters such as TopoGeo_AddPoint, TopoGeo_AddLineString, TopoGeo_AddPolygon, TopoGeo_LoadGeometry, and toTopoGeom use tolerance to snap incoming geometries to existing topology primitives.

Topology precision can also be inspected and repaired with ValidateTopologyPrecision and MakeTopologyPrecise, which operate on the topology precision grid or an explicit gridSize. This is closer to a metadata-backed precision model, but it is scoped to topology schemas and topological primitives rather than ordinary geometry values.

ST_AsMVTGeom has extent, buffer, and clip_geom parameters. These are not GEOS tolerance controls. The function transforms geometries into integer Mapbox Vector Tile coordinate space, snaps coordinates to the integer tile grid, removes repeated and collinear points, optionally clips to the tile bounds plus the tile-coordinate buffer, and may drop geometries smaller than half a tile-grid cell before deserializing.

For MVT output, simplify or reduce precision in map coordinates before calling ST_AsMVTGeom if the desired behavior is scale-dependent simplification rather than tile-coordinate quantization.

The old Trac page also asked whether PostGIS should infer precision from input coordinate decimals or carry a per-geometry/server-runtime precision model. Current code does not provide that metadata path. New work in this area should start as a design ticket that explains how it interacts with explicit gridSize, topology precision, geography behavior, MVT quantization, and existing tolerance parameters without changing predicate semantics silently.