Proximity Service (Uber)

Match riders to nearby drivers in real time using geohash grids and streaming location updates.

ExpertProduct14 min read

Read these first

Google Maps and Geospatial Systems

Why this matters

Ride-hailing, food delivery, and StreamHub's "join nearby watch party" all need sub-second nearest-neighbour lookup over millions of moving entities.
Brute-force distance calculation against every driver is O(n) per request — unacceptable at Uber scale (millions of drivers online).
Geohash grids or quad-trees partition space so you only scan a handful of cells, not the entire fleet.

Proximity service building blocks

Location ingestion — drivers push GPS updates via WebSocket or UDP every 2–4 seconds.
Spatial index — geohash prefix tree, Redis GEO, or custom quad-tree storing driver ID + coordinates.
Matching engine — finds top-K nearest available drivers, applies business rules (rating, vehicle type).
Dispatch queue — handles concurrent match requests; first-accept-wins or broadcast-to-top-3.
ETA service — routing engine estimates pickup time for each candidate.

High-level architecture

Proximity / ride-matching

Driver GPS → Kinesis → Redis GEO → match API → WebSocket to rider.

Java

POST /v1/location/update
Authorization: Bearer <driver_token>
Content-Type: application/json

{
  "driver_id": "drv_8f2a",
  "lat": 12.9716,
  "lng": 77.5946,
  "heading": 142,
  "timestamp_ms": 1740847200000
}

Java

POST /v1/rides/request
Content-Type: application/json

{
  "rider_id": "rdr_3c91",
  "pickup": { "lat": 12.9700, "lng": 77.5900 },
  "vehicle_type": "standard"
}

Geohash grid indexing

Geohash encodes lat/lng into a base-32 string. Shared prefixes mean geographic proximity.

Java

import geohash2

lat, lng = 12.9716, 77.5946
hash_6 = geohash2.encode(lat, lng, precision=6)  # "tdr1w3" — ~1.2 km cell
hash_7 = geohash2.encode(lat, lng, precision=7)  # "tdr1w3y" — ~150 m cell

Store drivers in Redis sorted sets keyed by geohash prefix:

Java

GEOADD drivers:tdr1w3 77.5946 12.9716 drv_8f2a
GEORADIUS drivers:tdr1w3 77.5900 12.9700 2 km ASC COUNT 10

Matching flow

Aspect	Strategy	Trade-off
Broadcast to top-3	Send ride offer to 3 nearest drivers	Fast match; may overload popular drivers
Serial dispatch	Offer to #1, wait 10s, then #2	Fairer; slower average match time
Batch matching	Optimise all pending rides every 5s	Better global assignment; higher latency
Surge pricing	Raise price in high-demand cells	Balances supply/demand dynamically

Broadcast to top-3
StrategySend ride offer to 3 nearest drivers
Trade-offFast match; may overload popular drivers
Serial dispatch
StrategyOffer to #1, wait 10s, then #2
Trade-offFairer; slower average match time
Batch matching
StrategyOptimise all pending rides every 5s
Trade-offBetter global assignment; higher latency
Surge pricing
StrategyRaise price in high-demand cells
Trade-offBalances supply/demand dynamically

Uber uses a mix: serial dispatch with surge multipliers per geohash cell.

Handling location staleness

Drivers go offline, lose signal, or stop updating. A location older than 30 seconds is unreliable.

Java

{
  "driver_id": "drv_8f2a",
  "lat": 12.9716,
  "lng": 77.5946,
  "updated_at": "2026-04-01T14:32:10Z",
  "status": "available",
  "ttl_seconds": 30
}

Background sweeper removes expired entries from the spatial index. On match, reject candidates whose updated_at exceeds the TTL.

Scale estimates

Metric	Estimate
Active drivers globally	5M peak
Location updates per second	5M / 3s ≈ 1.7M writes/s
Match requests per second	~50K peak
Geohash cells queried per match	9 (cell + neighbours)
Redis memory per driver	~100 bytes → 500 MB total

Active drivers globally
Estimate5M peak
Location updates per second
Estimate5M / 3s ≈ 1.7M writes/s
Match requests per second
Estimate~50K peak
Geohash cells queried per match
Estimate9 (cell + neighbours)
Redis memory per driver
Estimate~100 bytes → 500 MB total

Shard the spatial index by city or geohash prefix to keep each Redis cluster under 1M ops/s.

WebSocket connection management

Drivers maintain persistent WebSocket connections to regional gateway servers. On disconnect, mark driver unavailable immediately.

Java

# Connection gateway scaling (Kubernetes)
apiVersion: apps/v1
kind: Deployment
metadata:
  name: location-gateway
spec:
  replicas: 20
  template:
    spec:
      containers:
        - name: gateway
          resources:
            requests:
              cpu: "2"
              memory: 4Gi
          env:
            - name: MAX_CONNECTIONS_PER_POD
              value: "50000"

Quick recall

Everything you need if you only revisit this box.

Proximity = frequent location writes + fast spatial lookup, not full-table scans.
Geohash prefixes or Redis GEO partition space into queryable cells.
Match flow: query cell + neighbours → filter by status/TTL → rank by distance → dispatch.
Stale locations (>30s) must be evicted; disconnected drivers marked unavailable instantly.
Shard spatial indexes by region; 1.7M location writes/s requires horizontal partitioning.

Test yourself

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