Under the Hood: Why We Chose H3 Hexagons for Global Intelligence

In the world of location intelligence, the way you "bin" the world determines the quality of your insights. Traditional GIS has long relied on Census Tracts, ZIP codes, and rectangular grids to organize spatial data. But at Place Signals, as we scaled to process billions of data points across the globe, these traditional containers broke down.

To build a truly global, sub-second intelligence engine, we needed a mathematical foundation that didn't just store data, but accelerated it. This is why we chose Uber’s H3—a hexagonal, hierarchical spatial index—as the core of our architecture.

The Grid Problem: Beyond Squares and ZIPs

If you’ve ever worked with spatial data, you’ve likely encountered the Modifiable Areal Unit Problem (MAUP). ZIP codes and Census Tracts are political and administrative boundaries; they are irregular, change over time, and vary wildly in size. For a data scientist, they are a nightmare.

Rectangular grids (like Quadkeys) solve the irregularity but introduce directional bias. In a square grid, a cell has two types of neighbors: those sharing an edge (closer) and those sharing a vertex (further). This inconsistency creates "diagonal distortion" in distance-based calculations like trade area gravity or pedestrian accessibility.

Enter H3: The Hexagonal Advantage

Developed by Uber and open-sourced in 2018, H3 partitions the world into a seamless grid of hexagons. Unlike squares, hexagons offer three critical mathematical properties that make them the "Goldilocks" of spatial indexing.

1. Uniform Adjacency

Every hexagon has exactly six neighbors, and the distance from the center of a hexagon to the center of any of its neighbors is constant. This "perfect" adjacency eliminates the diagonal distortion found in square grids. When Place Signals calculates "Trade Area Gravity," the uniform distance between H3 cells ensures that our decay functions are mathematically consistent in every direction.

2. Hierarchical Efficiency

H3 isn't just one grid; it’s a system of 16 resolutions. A single H3 index encodes its parent and child relationships directly in its ID. This allows for extremely fast "up-sampling" and "down-sampling" via bitwise operations rather than expensive geometric intersections.

At Place Signals, we store high-fidelity demographics at Resolution 9 (roughly 0.1km²) and roll them up to Resolution 7 for regional heatmaps. Because these are hierarchical transitions, the computation is essentially "free."

3. Computational Speed

H3 represents every cell on Earth as a 64-bit integer. This is perhaps the most significant "under the hood" benefit. Traditional GIS operations like ST_Intersects require complex floating-point geometry math.

With H3, "Finding all amenities within a 1km radius" transforms from a geometric intersection into a disk-seek of pre-calculated neighbor indices. In our Core API, this enables us to perform spatial joins across millions of rows in under 50ms.

How H3 Powers Place Signals

We use H3 as the "Universal Join Key" for our entire data lake. Instead of joining messy school districts to noisy foot traffic polygons, we project every signal into the H3 grid:

• MetricHex Table: Our primary data structure, MetricHex, stores normalized indicators (Population, Disposable Income, Signal Vitality) keyed by H3 indices.
• Micro-Market Scores: By normalizing all data into a uniform H3 structure, we can calculate a "Vibrancy Score" for a specific street corner by simply summing the values of the surrounding hexes.
• Polyfill Aggregation: When a user draws a custom catchment area, we use h3.polyfill to identify the covering hexes and aggregate their pre-computed values instantly.

Conclusion

We chose H3 because high-stakes location intelligence requires the most efficient mathematical foundation on Earth. By moving away from arbitrary administrative lines and toward a rigid, hexagonal coordinate system, we’ve built a platform that is faster, more precise, and ready for global scale.

In the physical world, things are rarely square. In the data world, they should be hexagonal.