[HansVRP]

While working on the ML blueprint I figured it should not be hard to extend the spitting function towards point geometries as well.

@VincentVerelst you are quite familiar with this; can you please chekc and validate as well?

For me the results look good and respect tilegrids as well:

[HansVRP]

@VincentVerelst can you take a look?

[HansVRP]

@soxofaan could you have a look as well (I do still need to fix the unit tests)

[soxofaan]

can't go deep here, just some quick observations

[Copilot]

🟡 Changes recommended

Input features can be duplicated or lost, and custom tile grids do not satisfy the advertised interface contract.

3 open issues

• Reject empty geometries before spatial grouping · New
• Normalize duplicate GeoDataFrame indices before subdivision · New
• Support custom tile grid implementations · New

What changed in this PR

Adds point-based job splitting to the existing job-management utilities, supporting sample extraction workflows alongside area-based splitting.

Changes:

• Introduces split_points with optional feature limits, spatial subdivision, and tile-grid grouping.
• Exports the helper publicly and documents its usage.
• Adds tests for supported inputs, grouping, reprojection, and validation.

File	Description
tests/​extra/​job_management/​test_job_splitting.py	Tests point splitting and tile-grid behavior.
openeo/​extra/​job_management/​_job_splitting.py	Implements point grouping and recursive subdivision.
openeo/​extra/​job_management/​__init__.py	Exports split_points publicly.
docs/​cookbook/​job_manager.rst	Adds point-extraction batching examples.
CHANGELOG.md	Announces the new helper.

_{Review effort: Balanced}

[soxofaan]

FYI: any unit test failures you see about jsonschema are fixed with #960, so make sure to merge master to get rid of that problem

[Copilot]

🟡 Changes recommended

The change breaks existing split_area(aoi=...) keyword calls and uses a Shapely 2-only API that causes predefined-grid grouping to fail on supported Shapely 1.8 installations.

2 open issues

• Preserve the existing aoi keyword argument · New
• Support Shapely 1.8 in grid grouping · New

3 resolved since last review

• Normalize duplicate GeoDataFrame indices before subdivision
• Reject empty geometries before spatial grouping
• Support custom tile grid implementations

_{Review effort: Balanced}

[HansVRP]

Some additional info to the commit; I'll add some figures to illustrate the caveats

Grouping strategies:

• max_geometries: recursive median bisection along the longer axis of
  the representative points; groups stay spatially clustered with sizes
  between ~max/2 and max.

• tile_size + projection: features are binned per regular tile, then
  each tile is bisected. The grid is anchored at the CRS origin
  (floor(x / tile_size)), so adding an outlier does not change the
  grouping of existing features.

• tile_grid (e.g. Sentinel-2 MGRS): features are assigned to the tile
  containing their representative point; features outside all tiles are
  grouped separately. Without max_geometries: one group per tile.

  

Caveats (TODO):

• Groups are limited by feature count, not by extent unless a tile_gris is added. On 550 synthetic
  points (3 clusters + noise), the largest group extent is ~41 km
  without a grid and ~21 km with tile_size=20 km.

  

• Polygons are never split: large polygons extend beyond their tile, and
  groups are not balanced by area (use split_area for a single large AOI).

• Multi-part geometries stay in one group, placed by a single
  representative point. 40 MultiPolygon farms (3 distant parts each),
  max_geometries=10: 4 groups each spanning the whole AOI, versus
  16 compact groups (sizes 7..8) after explode(). The original index is
  kept, so exploded parts can be linked back to their rows.

• Overlapping tiles: the first matching tile in the grid wins (not the
  one with the largest overlap)

Input validation:

• Null or empty geometries raise JobSplittingFailure instead of being
  silently dropped.
• Duplicated index labels are supported: grouping is position-based and
  the original index is preserved.
• tile_grid cannot be combined with tile_size / projection.

[HansVRP]

@VincentVerelst @soxofaan ready for another look; I added a comment above with the caveats I know of.

[soxofaan]

short on time here, so only this small note

[soxofaan]

wrong place to put this changelog entry