[jravani]

-link-pylovo spent 29 minutes batching buildings before linking began, on one country database holding Deggendorf and Köln (27,420 buildings). The linking itself took under a second.

Cause

GetGridBatches built an ST_SquareGrid over the envelope of all buildings and joined every building's point to every cell. The two regions span 438 x 233 km, so 103,400 cells; the planner estimated 1,000 and chose a nested loop: about 2.8 billion ST_Intersects calls.

Nested Loop  (cost=3976.53..360242410.42 rows=27420)
  Join Filter: st_intersects(st_pointonsurface(b.geom), (st_squaregrid('1000', ...)).geom)
  ->  CTE Scan on all_buildings b  (rows=27420)
  ->  Materialize  (rows=1000)
        ->  ProjectSet ... st_squaregrid

Change

• Each building's cell is floor(x / size), floor(y / size) of a point inside its footprint, and batches are grouped by that index. The cells are the same origin-aligned squares, every building still lands in exactly one batch, and empty cells are never built.
• CHANGELOG.md: Fixes entry.

GroupAggregate (actual rows=49)
  ->  Gather Merge -> Sort -> Parallel Seq Scan on lod2_building (actual rows=9140, loops=3)
Execution Time: 38.983 ms

Tests

• Same database: 49 batches before and after, each of the 27,420 buildings in exactly one; 29 min before, 39 ms after.
• The five TestGetGridBatches_* integration tests and the rest of ./internal/process/ pass. No test fails before this change: the output is unchanged and only the cost differs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!