Google Maps System Design: Four Hard Problems

Open Google Maps and, in a second, you see a smooth zoomable map, a route avoiding traffic, and an arrival time that's usually right. Behind that are four separate systems:

  1. Rendering the map at any zoom level, anywhere on Earth
  2. Finding routes across a road graph with hundreds of millions of segments
  3. Knowing live traffic from millions of moving phones
  4. Predicting arrival times that account for traffic yet to come

This post explains each one, using what Google has published and the standard interview design.


๐ŸŽฏ Quick Answer (30-Second Read)

  • Map tiles: the world is pre-cut into square tiles at each zoom level; apps download only visible tiles (now mostly vector tiles).
  • Routing: roads form a graph; algorithms like Dijkstra/A* plus precomputed shortcuts make long routes fast.
  • Traffic: anonymised, aggregated location data from phones shows how fast roads are moving.
  • ETA: Google and DeepMind described using graph neural networks to predict travel times, improving accuracy in cities worldwide.
  • Search: places, addresses and businesses come from a huge geospatial index.

High-Level Architecture

flowchart TB APP["๐Ÿ“ฑ Maps app"] --> TILE["Tile service"] TILE --> CDN["CDN / edge cache
(popular tiles)"] CDN --> TS[("Pre-rendered vector tiles
per zoom level")] APP --> SEARCH["Places search"] SEARCH --> PI[("Place index
businesses, addresses")] APP --> ROUTE["Routing service"] ROUTE --> RG[("Road graph
+ precomputed shortcuts")] ROUTE --> ETA["ETA model
(graph neural network)"] PHONES["๐Ÿ“ Phones sharing location
(opt-in, anonymised)"] --> TR["Traffic pipeline"] TR --> LIVE[("Live road speeds")] LIVE --> ROUTE LIVE --> ETA

1. Map Tiles: Cutting the World Into Squares

At zoom level 0, the whole world fits in one tile. Each zoom level splits every tile into four:

Zoom level Tiles to cover the world
0 1
1 4
5 1,024
10 ~1 million
15 ~1 billion

Formula: 4^z tiles at zoom z. The app requests only the few dozen tiles visible on screen. Popular tiles (cities) are cached at the edge, so most requests never reach Google's core servers.

Vector vs raster: older maps sent images (raster tiles). Modern apps get vector tiles (roads, shapes and labels as data) and draw them on the device, which enables smooth zooming, rotation, 3D and smaller downloads.


2. Routing: Shortest Paths on a Huge Graph

Intersections are nodes, road segments are edges, and each edge has a cost (travel time).

flowchart LR A(("Home")) -->|"4 min"| B(("Junction")) A -->|"9 min"| C(("Flyover")) B -->|"12 min (traffic)"| D(("Office")) C -->|"5 min"| D

Here, Home โ†’ Flyover โ†’ Office (14 min) beats Home โ†’ Junction โ†’ Office (16 min), even though the first leg is longer.

  • Dijkstra's algorithm finds shortest paths but explores too many roads for a cross-country route.
  • A* focuses the search towards the destination.
  • Preprocessing (like contraction hierarchies) adds shortcut edges so long routes jump across highways instead of exploring every street.
Key takeaway: fast routing comes from doing heavy work ahead of time. Precomputed shortcuts let a route across India be found in milliseconds instead of checking millions of streets.

3. Live Traffic From Phones

When people opt in to share location, their phones send anonymised speed and position data. Aggregating many devices per road segment reveals traffic: if many phones move at 8 km/h on a road normally driven at 40 km/h, it's jammed. Google combines this with historical patterns, and with user reports of incidents.


4. ETA Prediction

An ETA has to predict future traffic along the route. Google and DeepMind published work in 2020 describing how they used graph neural networks that model road segments as a connected graph, improving ETA accuracy significantly in cities including Taipei, Sydney and others.


5. Places Search

"Chai near me" needs a geospatial index plus text search and ranking (distance, ratings, open now, relevance). Like Uber's map cells, Maps uses spatial cell systems; Google's open-source S2 library divides the sphere into hierarchical cells. (See Uber System Design.)


Interview Checklist

  • Tile pyramid (4^z), vector tiles, CDN caching
  • Road graph with preprocessing for fast routing
  • Traffic pipeline from aggregated location data
  • ETA model using historical + live data
  • Geospatial search with ranking
  • Offline maps: download tiles and routing data for a region

Frequently Asked Questions

How does Google Maps know about traffic?

Google Maps aggregates anonymised location and speed data from phones that have location sharing enabled, combined with historical traffic patterns and user incident reports. When many devices move slowly on a road segment, Maps shows it as congested.

What algorithm does Google Maps use for directions?

Google doesn't publish its exact routing system, but route planning on large road networks typically uses shortest-path algorithms like Dijkstra or A* combined with preprocessing such as contraction hierarchies, plus live and predicted traffic as edge costs.

How does Google Maps predict arrival time?

It combines live traffic, historical patterns for that time and day, and machine-learning models. Google and DeepMind described using graph neural networks that model connected road segments, which improved ETA accuracy in many cities.


Conclusion

Google Maps combines a tile pyramid served from CDNs for the map, a preprocessed road graph for fast routing, aggregated phone data for live traffic, and machine-learning models for ETAs. Each part does heavy work in advance or in aggregate so the answer on your phone appears in about a second.


Sources

Last verified: 2 November 2026.


Related reads: How Google Street View Images Were First Captured ยท Uber System Design ยท How Swiggy Built Logistics Without Addresses ยท YouTube System Design

I write about system design and how real products are built. Follow on LinkedIn or X, or invite me to speak.