Files
ArchiTools/geoportal/TILE-SERVER-EVALUATION.md
AI Assistant ee86af6183 docs: update tile evaluation + monitoring + add geoportal improvement mega prompt
- TILE-SERVER-EVALUATION.md: updated to reflect current architecture (PMTiles z0-z18)
- MODULE-MAP.md: added PMTiles + tile-cache to Geoportal section
- Monitor: timeout increased to 90 min for z18 builds, description updated
- Added PROMPT-GEOPORTAL-IMPROVE.md with mega prompt for future sessions
  (includes MLT check, mvt-rs evaluation prompt, operational commands)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-28 18:38:53 +02:00

13 KiB

Tile Server Evaluation — ArchiTools Geoportal (March 2026)

Context

ArchiTools Geoportal serves vector tiles (MVT) from PostgreSQL 16 + PostGIS 3 via Martin. Data: ~330K GIS features (parcels, buildings, admin boundaries) in EPSG:3844 (Stereo70), growing to 1M+. Frontend: MapLibre GL JS 5.21, Next.js 16, Docker self-hosted via Portainer CE.


Problem Statement

  1. Martin v0.15.0 was running in auto-discovery mode — the existing martin.yaml config was never mounted
  2. Building labels (cadastral_ref) missing from MVT tiles despite the view exposing them
  3. Performance concerns at scale (330K -> 1M+ features)

Solutions Evaluated (7 options + emerging tech)

1. Martin (Fix + Upgrade) — WINNER

Aspect Detail
Root cause martin.yaml not mounted in docker-compose — Martin ran in auto-discovery mode
Fix Bake config into custom image via Dockerfile + upgrade v0.15 -> v1.4.0
Performance Fastest tile server benchmarked (2-3x faster than #2 Tegola)
EPSG:3844 Native support via default_srid: 3844
New in v1.4 ZSTD compression, MLT format, materialized views, better logging

Status: IMPLEMENTED AND VERIFIED IN PRODUCTION (2026-03-27)

2. pg_tileserv (CrunchyData)

Aspect Detail
Architecture Go binary, zero-config, delegates ST_AsMVT to PostGIS
Property control Auto from schema + URL ?properties= parameter
Performance 2-3x slower than Martin (Rechsteiner benchmark)
EPSG:3844 Supported (auto-reprojects via ST_Transform)
Killer feature Function-based sources (full SQL tile functions)
Dealbreaker View extent estimation bug (#156) affects all our views, development stagnant (last release Feb 2025)

Verdict: NO — slower, buggy with views, stagnant development.

3. Tegola (Go-based)

Aspect Detail
Architecture Go, TOML config, explicit per-layer SQL
Performance 2nd in benchmarks, but 2-3x slower than Martin
Built-in cache File, S3/MinIO, Redis — with seed/purge CLI
EPSG:3844 NOT SUPPORTED (only 3857/4326) — requires ST_Transform in every query
Killer feature Built-in tile seeding and cache purging

Verdict: NO — EPSG:3844 not supported, dealbreaker for our data.

4. t-rex (Rust-based)

Aspect Detail
Status Abandoned/unmaintained — no releases since 2023

Verdict: NO — dead project.

5. GeoJSON Direct from Next.js API

Aspect Detail
330K features ~270 MB uncompressed, 800 MB-1.4 GB browser memory
Browser impact 10-30s main thread freeze, mobile crash
Pan/zoom Full re-fetch on every viewport change, flickering
Viable range Only at zoom 16+ with <500 features in viewport

Verdict: NO — does not scale beyond ~20K features.

6. PMTiles (Pre-generated)

Aspect Detail
Architecture Single-file tile archive, HTTP Range Requests, no server needed
Performance ~5ms per tile (vs 200-2000ms for Martin on low-zoom)
Property control tippecanoe gives explicit include/exclude per property
Update strategy Full rebuild required (~3-7 min for 330K features)
EPSG:3844 Requires reprojection to 4326 via ogr2ogr before tippecanoe
MinIO serving Yes — direct HTTP Range Requests with CORS

Verdict: YES as hybrid complement — excellent for static UAT overview layers (z0-z12), Martin for live detail.

7. Emerging Solutions

Solution Status Relevance
mvt-rs (Rust) v0.16.2, active Admin UI, auth per layer, cache — good for multi-tenant
MLT format Stable Jan 2026 6x compression, 4x faster decode — Martin v1.3+ supports it
BBOX Maturing Similar to Tegola performance, unified raster+vector
DuckDB tiles Early Not PostGIS replacement, interesting for GeoParquet
FlatGeobuf Stable Good for <100K features, not a tile replacement

Benchmark Reference (Rechsteiner, April 2025)

Rank Server Language Relative Speed
1 Martin Rust 1x (fastest)
2 Tegola Go 2-3x slower
3 BBOX Rust ~same as Tegola
4 pg_tileserv Go ~4x slower
5 TiPg Python Slower
6 ldproxy Java 4-70x slower

Source: github.com/FabianRechsteiner/vector-tiles-benchmark


Implementation Roadmap

Phase 1: Martin Fix — DONE (2026-03-27)

Changes applied:

  • martin.Dockerfile: custom image that COPY-s martin.yaml into /config/
  • docker-compose.yml: Martin v0.15 -> v1.4.0, build from Dockerfile, --config flag
  • martin.yaml: comment updated to reflect v1.4
  • map-viewer.tsx: building labels layer activated (cladiriLabel at minzoom 16)

Deployment Lessons Learned

  1. Docker image tag format changed at v1.0: old tags use v prefix (v0.15.0), new tags do not (1.4.0). The tag ghcr.io/maplibre/martin:v1.4.0 does NOT exist — correct is ghcr.io/maplibre/martin:1.4.0.

  2. Portainer CE volume mount pitfall: volume ./martin.yaml:/config/martin.yaml:ro fails because Portainer deploys only the docker-compose.yml content, not the full git repo. Docker silently creates an empty directory instead of failing. Solution: bake config into a custom image with a 2-line Dockerfile:

    FROM ghcr.io/maplibre/martin:1.4.0
    COPY martin.yaml /config/martin.yaml
    
  3. Martin config format is stable: YAML format unchanged from v0.15 to v1.4 — postgres.tables, connection_string, auto_publish, properties map all work identically. No migration needed.

  4. PostGIS view geometry type: Martin logs UNKNOWN GEOMETRY TYPE for all views — this is normal for nested views (SELECT * FROM parent_view). Views don't register specific geometry types in geometry_columns. Does not affect tile generation or property inclusion.

Phase 2A: nginx Tile Cache — DONE (2026-03-27)

Impact: 10-100x faster on repeat requests, zero PostGIS load for cached tiles.

Changes applied:

  • nginx/tile-cache.conf: proxy_cache config with 2GB cache zone, 7-day TTL, stale serving
  • tile-cache.Dockerfile: bakes nginx config into custom image (Portainer CE pattern)
  • docker-compose.yml: tile-cache container, Martin no longer exposed on host
  • Gzip passthrough (Martin already compresses), browser caching via Cache-Control headers
  • CORS headers for cross-origin tile requests

Phase 2B: PMTiles — DONE (2026-03-27)

Impact: Sub-10ms overview tiles, zero PostGIS load for z0-z18.

Changes applied:

  • scripts/rebuild-overview-tiles.sh: ogr2ogr export (3844->4326) + tippecanoe generation
  • PMTiles archive: z0-z18, ~1-2 GB, includes all terenuri, cladiri, UATs, and administrativ layers
  • map-viewer.tsx: pmtiles:// protocol registered on MapLibre, hybrid source switching
  • MinIO bucket tiles with public read + CORS for Range Requests
  • N8N webhook trigger for rebuild (via monitor page)
  • Monitor page (/monitor): rebuild + warm-cache actions with live status polling

Phase 2C: MLT Format — DEFERRED

Martin v1.4 advertises MLT support, but it cannot generate MLT from PostGIS live queries. MLT generation requires pre-built tile archives (tippecanoe does not output MLT either). No actionable path until Martin or tippecanoe adds MLT output from PostGIS sources.

Phase 2D: mvt-rs Evaluation — FUTURE (Multi-Tenant)

Impact: Built-in auth, admin UI, per-layer access control. Effort: 1-2 days for evaluation + migration.

Reserved for when external client access to the geoportal is needed. mvt-rs (v0.16.2+, Rust, Salvo framework) provides per-layer auth and admin UI.


Phase 3: Current Architecture (as of 2026-03-27)

Full tile-serving pipeline in production:

PostGIS (EPSG:3844)
  |
  +--> Martin v1.4.0 (live MVT from 9 PostGIS views)
  |       |
  |       +--> tile-cache (nginx reverse proxy, 2GB disk, 7d TTL)
  |               |
  |               +--> Traefik (tools.beletage.ro/tiles)
  |
  +--> ogr2ogr (3844->4326) + tippecanoe (z0-z18)
          |
          +--> PMTiles archive (~1-2 GB)
                  |
                  +--> MinIO bucket "tiles" (HTTP Range Requests)
                          |
                          +--> MapLibre (pmtiles:// protocol)

Hybrid strategy:

  • PMTiles serves pre-generated overview tiles (all zoom levels, all layers)
  • Martin serves live detail tiles (real-time PostGIS data)
  • nginx tile-cache sits in front of Martin to absorb repeat requests
  • Rebuild triggered via N8N webhook from the /monitor page

Operational Commands

Rebuild PMTiles

Trigger from the Monitor page (/monitor -> "Rebuild PMTiles" button), which sends a webhook to N8N. N8N runs scripts/rebuild-overview-tiles.sh on the server.

Manual rebuild (SSH to 10.10.10.166):

cd /path/to/architools
bash scripts/rebuild-overview-tiles.sh

Warm nginx Cache

Trigger from the Monitor page (/monitor -> "Warm Cache" button). Pre-loads frequently accessed tiles into the nginx disk cache.

Purge nginx Tile Cache

docker exec tile-cache rm -rf /var/cache/nginx/tiles/*
docker exec tile-cache nginx -s reload

Restart Martin (after PostGIS view changes)

docker restart martin

Martin caches source schema at startup — must restart after DDL changes to pick up new columns.

Check PMTiles Status

# Check file size and last modified in MinIO
docker exec minio mc stat local/tiles/overview.pmtiles

Key Technical Details

Martin v1.4.0 Deployment Architecture

Gitea repo (martin.yaml + martin.Dockerfile)
  -> Portainer CE builds custom image: FROM martin:1.4.0, COPY martin.yaml
  -> Container starts with --config /config/martin.yaml
  -> Reads DATABASE_URL from environment
  -> Serves 9 PostGIS view sources on port 3000
  -> Host maps 3010:3000
  -> Traefik proxies tools.beletage.ro/tiles -> host:3010

Critical: Do NOT use volume mounts for config files in Portainer CE stacks. Always bake configs into custom images via Dockerfile COPY.

Martin Config (validated compatible v0.15 through v1.4)

The martin.yaml at project root defines 9 sources with explicit properties. Config format unchanged from v0.15 to v1.4 — no migration needed.

Key config features used:

  • auto_publish: false — only explicitly listed sources are served
  • default_srid: 3844 — all sources use Stereo70
  • properties: map per source — explicit column name + PostgreSQL type
  • minzoom/maxzoom per source — controls tile generation range
  • bounds: [20.2, 43.5, 30.0, 48.3] — approximate Romania extent

Docker Image Tag Convention

Martin changed tag format at v1.0:

  • Pre-1.0: ghcr.io/maplibre/martin:v0.15.0 (with v prefix)
  • Post-1.0: ghcr.io/maplibre/martin:1.4.0 (no v prefix)
  • Also available: latest, nightly

PostGIS View Chain

GisFeature table (Prisma) -> gis_features view -> gis_terenuri / gis_cladiri / gis_administrativ
                                                -> gis_terenuri_status / gis_cladiri_status (with JOINs)
GisUat table -> gis_uats_z0/z5/z8/z12 (with ST_SimplifyPreserveTopology)

MapLibre Layer Architecture

Sources (Martin):     gis_uats_z0, z5, z8, z12, administrativ, terenuri, cladiri
Layers per source:    fill + line + label (where applicable)
Selection:            Separate highlight layers on terenuri source
Drawing:              GeoJSON source for freehand/rect polygon
Building labels:      cladiriLabel layer, cadastral_ref at minzoom 16

Deployment Pitfalls (Discovered During Implementation)

  1. Portainer CE does not expose repo files to containers at runtime. Volume mounts like ./file.conf:/etc/file.conf:ro fail silently — Docker creates an empty directory. Always bake config files into custom images via Dockerfile COPY.

  2. Martin Docker tag format change at v1.0. v1.4.0 does not exist, 1.4.0 does. Always check ghcr.io/maplibre/martin for actual tags.

  3. Martin logs UNKNOWN GEOMETRY TYPE for PostGIS views. This is normal — nested views don't register geometry types in geometry_columns. Does not affect functionality.

  4. Martin auto-discovery mode is unreliable for property inclusion. Always use explicit config with auto_publish: false and per-source properties: definitions.

  5. Martin caches source schema at startup. After PostGIS view DDL changes (e.g., adding columns to gis_features), Martin must be restarted to pick up new columns.


References