T-Rex doesn't handle invalid geometry results correctly when transforming between coordinate systems

[hamarituc]

I encountered an issue where invalid point geometries are served.

Steps to reproduce

1. Consider the following data in a PostGIS database

BEGIN;
CREATE TABLE "public"."test" ( "ogc_fid" SERIAL, CONSTRAINT "test_pk" PRIMARY KEY ("ogc_fid") );
SELECT AddGeometryColumn('public','test','wkb_geometry',25833,'POLYGON',2);
CREATE INDEX "test_wkb_geometry_geom_idx" ON "public"."test" USING GIST ("wkb_geometry");
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E964000001000000050000009AA6C935869C154110164AAEEB7D55410E92AED1A09C15410E9FB145EE7D554178EADB7FCD9C15413B2A246DEB7D55418E44E7E2989C15411D3A977BEA7D55419AA6C935869C154110164AAEEB7D5541', 1);
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E9640000010000000500000092E008F4179D1541534A8CE8EF7D5541BA518FE92F9D1541A47BCA5FED7D5541341B6238FA9C154171581624EC7D5541C3CA4C05DF9C15413675F754F07D554192E008F4179D1541534A8CE8EF7D5541', 2);
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E96400000100000006000000023A12A91A9D154167CE46A0F37D55413236D5033A9D15413FFE54E8F27D5541DCBE1A20359D1541CABD12D9F17D55413E4F782A419D1541078DD709F07D554118FCCDC60E9D154136BDE349F07D5541023A12A91A9D154167CE46A0F37D5541', 3);
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E9640000010000000400000023173BB3B89C1541FD196035EA7D55415954A095C79C1541A63D3AA9E87D5541AB142256D79C154159990D34EA7D554123173BB3B89C1541FD196035EA7D5541', 4);
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E964000001000000050000002EE515095A9C15417FC68947F47D5541704A76218B9C1541618A4738F77D5541DFE97506CC9C1541E1C4BEF9F17D5541133529E45A9C15413048178FF17D55412EE515095A9C15417FC68947F47D5541', 5);
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E964000001000000050000005C84D2275C9C15418E1EA353F07D554176BCCBA6819C15411D415EFCEF7D55410C1A04257F9C1541F43B6077EE7D55414C1200FB4C9C1541E9A6B3E9EE7D55415C84D2275C9C15418E1EA353F07D5541', 6);
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E9640000010000000500000055487DCD589C154156222501EE7D5541C714C0706E9C154184D3A3C3ED7D554156540F1E6C9C1541D7E2667EEC7D55412966C7475C9C15415A164C7FEC7D554155487DCD589C154156222501EE7D5541', 7);
INSERT INTO "public"."test" ("wkb_geometry" , "ogc_fid" ) VALUES ('0103000020E96400000100000005000000147B7306D69C154124585008F67D55413EC200D2D99C1541356B1CEBF27D55413D5F9FE8FC9C1541397C7E6BF27D554106D8B1FEFA9C1541043E24ADF57D5541147B7306D69C154124585008F67D5541', 8);
COMMIT;

2. Configure T-Rex to serve polygon centroids from this database

[service.mvt]
viewer = false

[[datasource]]
name = "test"
dbconn = "as required"

[grid]
predefined = "web_mercator"

[[tileset]]
name = "test"
#minzoom = 15
#maxzoom = 22
extent = { minx = 12.9053, maxx = 12.9609, miny = 50.7920, maxy = 50.8538 }
center = [ 12.92804, 50.83904 ]
start_zoom = 17

[[tileset.layer]]
name = "centroid"
datasource = "test"
geometry_field = "geom"
geometry_type = "POINT"
srid = 25833
fid_field = "ogc_fid"
# Clip polygons with a buffer
buffer_size = 200
simplify = false

  [[tileset.layer.query]]
  minzoom = 17
  maxzoom = 24
  sql = "SELECT ogc_fid, ST_PointOnSurface(wkb_geometry) AS geom FROM test WHERE ST_PointOnSurface(wkb_geometry) && !bbox!"

[cache.file]
base = "/tmp/trex/cache"

[webserver]
bind = "localhost"
port = 6767

3. Fetch a tile from /vector/test/19/280971/175891.pbf

Observed Behavior

Loading 175891.pbf into QGis reveals there is a single point at the lower left corner of the tile not present in the database:

Expected Behavior

The served tile should not contain invalid geometries.

I suppose this invalid geometries arise when the databases uses a different coordinate reference system than Web mercator. According to the docs the &&-operator should be used inside the WHERE-clause to filter for all relevant geometries for the current tile. This operator compares the bounding boxes of both of its operands. Later on T-Rex clips the result of the query by means of the ST_Intersection()-function to the exact tile extent.

SELECT
  ST_Transform(
    ST_Intersection(
      geom,
      ST_Transform(
        ST_Segmentize(
          ST_MakeEnvelope(
            1439079.9315250143-12.5*0.2985821417389698::FLOAT8,
            6592846.563653022-12.5*0.2985821417389698::FLOAT8,
            1439156.3685532995+12.5*0.2985821417389698::FLOAT8,
            6592923.000681307+12.5*0.2985821417389698::FLOAT8,
            3857
          ),
          (1439156.3685532995-1439079.9315250143)/512
        ),
        25833
      )
    ),
    3857
  )
AS geom,"ogc_fid"
FROM
(
  SELECT ogc_fid, ST_PointOnSurface(wkb_geometry) AS geom
  FROM test
  WHERE ST_PointOnSurface(wkb_geometry) &&
    ST_Transform(
      ST_Segmentize(
        ST_MakeEnvelope(
          1439079.9315250143-12.5*0.2985821417389698::FLOAT8,
          6592846.563653022-12.5*0.2985821417389698::FLOAT8,
          1439156.3685532995+12.5*0.2985821417389698::FLOAT8,
          6592923.000681307+12.5*0.2985821417389698::FLOAT8,
          3857
        ),
        (1439156.3685532995-1439079.9315250143)/512
      ),
      25833
    )
) AS _q

The tile area (including a buffer, if configured) itself is transformed into the database native CRS (light-red polygon). The dark-red box shows the bounding box of the tile area. The green centroid therefore fulfills the WHERE-clause, but is subsequently clipped in the SELECT clause which yields an POINT EMPTY value for this row.

T-Rex doesn't handle that case correctly. It should ignore empty point geometries in this case.

[pka]

I've reproduced this case with bbox-tile-server and it looks better:

This was expected, since null geometries are explicetely ignored in the PostGIS driver of bbox-tile-server.

[hamarituc]

Is there any chance for a native debian package of bbox? I would like to check it out for my setup.

[pka]

This asset includes a Debian package: https://github.com/bbox-services/bbox/actions/runs/8375677003/artifacts/1346593260

You can run it with bbox-tile-server --t-rex-config=t_rex.toml serve

[hamarituc]

This asset includes a Debian package: https://github.com/bbox-services/bbox/actions/runs/8375677003/artifacts/1346593260

You can run it with bbox-tile-server --t-rex-config=t_rex.toml serve

Thanks for providing a Debian package so quickly. It doesn't install on Debian 12. Is there a reason for the dependency on Glibc 2.34 or does 2.31 work anyway?

# LANG="" dpkg -i bbox-tile-server_0.5.0~beta2_amd64.deb
(Reading database ... 111720 files and directories currently installed.)
Preparing to unpack bbox-tile-server_0.5.0~beta2_amd64.deb ...
Unpacking bbox-tile-server (0.5.0~beta2) over (0.5.0~beta2) ...
dpkg: dependency problems prevent configuration of bbox-tile-server:
 bbox-tile-server depends on libc6 (>= 2.34); however:
  Version of libc6:amd64 on system is 2.31-13+deb11u8.

dpkg: error processing package bbox-tile-server (--install):
 dependency problems - leaving unconfigured
Errors were encountered while processing:
 bbox-tile-server

[pka]

The package was built on Ubuntu 22.04. Will do an other one on Debian 12 then.

[pka]

There's a bookworm package in the latest CI build: https://github.com/bbox-services/bbox/actions/runs/8392289824/artifacts/1350419108

[hamarituc]

It still requires glibc 2.34.

# LANG="" dpkg -i bbox-tile-server_0.5.0~beta2-bookworm_amd64.deb
Selecting previously unselected package bbox-tile-server.
(Reading database ... 111716 files and directories currently installed.)
Preparing to unpack bbox-tile-server_0.5.0~beta2-bookworm_amd64.deb ...
Unpacking bbox-tile-server (0.5.0~beta2-bookworm) ...
dpkg: dependency problems prevent configuration of bbox-tile-server:
 bbox-tile-server depends on libc6 (>= 2.34); however:
  Version of libc6:amd64 on system is 2.31-13+deb11u8.

dpkg: error processing package bbox-tile-server (--install):
 dependency problems - leaving unconfigured
Errors were encountered while processing:
 bbox-tile-server

May it be worth of transfering this issue to the bbox repository as it is not related to the original issue anymore?

[pka]

According to the error message you seem to be on bullseye not bookworm. Made another package to try: https://github.com/bbox-services/bbox/actions/runs/8422167915/artifacts/1356021483

[hamarituc]

According to the error message you seem to be on bullseye not bookworm. Made another package to try: https://github.com/bbox-services/bbox/actions/runs/8422167915/artifacts/1356021483

Sorry for this case of PEBKAC. It was still too early in the morning. I tried the package on my production server running T-Rex under Debian 11 instead on my development server running Debian 12. The package runs well under Debian 12. Thank you.

Hope you will provide Debian Packages as a official release.

Many thanks for your work.