What geospatial forbids

You cannot partition extended objects disjointly without duplicating them.
You cannot linearize two dimensions without tearing neighbourhoods.
You cannot project a sphere onto a plane preserving area, angle and distance together.
You cannot decide degenerate predicates reliably in floating point.
You cannot estimate spatial query cost from cardinality alone, because the filter's selectivity depends on the shapes.

The escapes and their prices

Escape	Price
Bounding boxes give near-free rejection	false positives proportional to shape irregularity
Duplication across cells gives disjoint partitions and clean pruning	index growth and read-side deduplication
Overlapping regions give single storage	multi-path search, which degrades toward a scan under heavy overlap
Grid cell ids give the whole existing relational toolchain	tear artefacts, a resolution knob, coarser answers
Exact arithmetic gives robust predicates	cost, so it is used only for degenerate cases
Adaptive partitioning gives balance	being fitted to a distribution that will move

Case, resolvedConstructed case · Amara at Haatly

A courier 10 metres away, never offered the job

A courier 10 metres away across a tile boundary was never offered a job, a zone straddling two partitions was counted twice, and two libraries disagreed on a point on a zone's edge. Where it started: Geospatial →

Report 1 (the missed courier). The index grid cut the world into cells. The 500 m circle around the pickup crossed a cell boundary, but the query looked only in the pickup's own cell. Amara changes the query to cover the circle with cells: the pickup's cell plus the neighbouring cells the circle touches. She deliberately over-covers (a slightly larger set of cells), then refines with an exact distance on the spheroid in metres. Over-covering gives false positives that the exact check removes. Under-covering would lose couriers for good.

Report 2 (double counting). The zone polygon was stored in both partitions, because duplication is the price of disjoint cells. The weekly report now deduplicates by zone id before summing. Revenue per zone drops back to the true figure.

Report 3 (inside or outside). She enforces an SRID on every geometry, validates polygons at write time, and fixes one library and one predicate (ST_Covers rather than a mixture) for the "on the edge" case, so the convention is declared and reproducible. The truth about a point exactly on an edge is not recoverable from floating point, so a declared rule is the only fix.

She also stores the zone's bounding box and an inner box in the index so that most pickups are accepted or rejected without exact geometry.