Google Maps and Geospatial Systems

Tile rendering, routing graphs and geospatial indexes for location-aware applications.

ExpertProduct15 min read

Read these first

Hotel Reservation System

Why this matters

Location features appear in ride-hailing, delivery, social check-ins, and StreamHub's live-event discovery — all share the same geospatial primitives.
Naive latitude/longitude B-tree queries degrade at scale; purpose-built spatial indexes (R-tree, geohash grids, H3) are the difference between 5 ms and 5 s.
Tile pyramids and vector tiles let you serve the entire planet without loading one giant image per request.

Core components of a maps platform

Tile server — pre-rendered raster or vector tiles at zoom levels 0–20; clients request only visible tiles.
Routing engine — directed graph of road segments with weights (distance, time, tolls); A* or Contraction Hierarchies for fast pathfinding.
Geospatial index — R-tree, geohash prefix, or H3 hex cells for proximity and bounding-box queries.
Reverse geocoder — converts coordinates to human-readable addresses using a spatial + text index.
Traffic overlay — real-time speed data merged into routing weights via streaming ingestion.

Tile pyramid architecture

Geospatial maps architecture

CloudFront tile CDN; PostGIS RDS routing; cached traffic API.

Clients request map tiles by zoom level and tile coordinates (z, x, y). At zoom 0 the world is one tile; each level quadruples tile count.

Java

GET /tiles/14/4821/6156.pbf
Host: tiles.streamhub.com
Accept: application/vnd.mapbox-vector-tile

Routing graph design

Road networks are stored as a graph: nodes are intersections, edges are road segments with attributes.

Java

{
  "edge_id": "e_88421",
  "from_node": 10442,
  "to_node": 10443,
  "length_m": 312,
  "speed_limit_kmh": 50,
  "one_way": true,
  "geometry": [[12.97, 77.59], [12.971, 77.591]]
}

Aspect	Algorithm	When to use
Dijkstra / A*	Correct shortest path	Small graphs, offline preprocessing
Contraction Hierarchies	Preprocessed shortcuts	Production routing at Google/OSM scale
Isochrone queries	Reachability within N minutes	Delivery radius, EV range
Multi-modal	Walk + transit + drive legs	Urban mobility apps

Dijkstra / A*
AlgorithmCorrect shortest path
When to useSmall graphs, offline preprocessing
Contraction Hierarchies
AlgorithmPreprocessed shortcuts
When to useProduction routing at Google/OSM scale
Isochrone queries
AlgorithmReachability within N minutes
When to useDelivery radius, EV range
Multi-modal
AlgorithmWalk + transit + drive legs
When to useUrban mobility apps

Routing latency targets: under 100 ms for a single origin-destination query at production scale.

Geospatial indexing for POIs

Points of interest (restaurants, venues, parking) live in a spatial index keyed by location.

Java

-- PostGIS example: find venues within 2 km of a point
SELECT id, name, ST_Distance(location, ST_MakePoint(77.59, 12.97)::geography) AS dist_m
FROM venues
WHERE ST_DWithin(location, ST_MakePoint(77.59, 12.97)::geography, 2000)
ORDER BY dist_m
LIMIT 20;

For StreamHub event discovery, index live-stream venues by H3 resolution-7 cells (~1.2 km²) so "events near me" is a cell lookup plus a small refinement query.

Scale estimates for StreamHub maps

Metric	Estimate
DAU using map features	5M
Tile requests per user session	~40
Peak tile QPS	5M × 40 / 86400 × 3 ≈ 7K
Routing queries per day	500K
POI database size	~200M points globally

DAU using map features
Estimate5M
Tile requests per user session
Estimate~40
Peak tile QPS
Estimate5M × 40 / 86400 × 3 ≈ 7K
Routing queries per day
Estimate500K
POI database size
Estimate~200M points globally

Back-of-envelope for a mid-scale maps layer — CDN caches 95%+ of tile traffic.

Failure modes and mitigations

Production concerns

Stale tiles — version tiles by etag or timestamp; invalidate CDN cache on map data updates.
Routing divergence — A/B test new graph versions in shadow mode before promoting to production.
Hot regions — pre-warm CDN cache for major events; scale tile servers horizontally per region.
GPS inaccuracy — snap user coordinates to nearest road segment before routing.

Quick recall

Everything you need if you only revisit this box.

Maps = tile pyramid (imagery) + routing graph (directions) + spatial index (nearby search).
Vector tiles (.pbf) are smaller and styleable; raster tiles are simpler but less flexible.
Use PostGIS R-tree, geohash prefixes, or H3 for proximity — not naive lat/lng range scans.
Contraction Hierarchies make continent-scale routing sub-100 ms.
Cache tiles at CDN edge; keep routing and POI queries on regional clusters.

Test yourself

Answer these before moving on — recall is what makes it stick.