Geospatial Situation Detection through FlinkCEP @Uber
summary
Uber runs situation detection across 700+ cities using Apache Flink and FlinkCEP to match temporal sequences, with spatial clustering implemented as DBSCAN on H3 hex tiles (H3 cells down to ~10 sq ft). The pipeline produces changelogs of incremental cluster adds/merges/splits, supports backtesting via event-time semantics, smooths load with per-partition random window offsets to avoid thundering‑herd spikes, and layers online calibration on offline ML models for non‑stationary traffic.
AI-written summary · about
chapters (16)
- 0:00Geospatial Situation Detection through FlinkCEP @Uber
- 1:10Marketplace
- 3:11Observability Problem
- 4:14Scaling Observability
- 7:58Detecting the Region through clustering
- 8:50Density-based Spatial Clustering with Noise
- 10:48H3: Hexagonify the World!
- 11:41Uniform Adjacency
- 14:09Low-Latency Clustering on Streams
- 16:53Clustering on streaming data
- 19:46Complex Event Processing
- 21:03CEP: Cluster as Primitive
- 23:08Human in the loop
- 24:00Dynamic Rule Configuration
- 24:48Staggering Window
- 27:22Continuous Learning on Streams
related talks
40mStreaming Geospatial Vector Data with Apache Projects
19mExplore geocoded datasets with PwC’s innovative data platform
11mUrban Computing
26mUber’s Intelligent Insights Assistant
30mBuilding geospatial apps on Postgres, PostGIS, & Citus at large scale
48mData Streaming with Apache Beam / Cloud Dataflow
06mLightning Talk: Leveraging Realtime Geospatial Data and the Spring Ecosystem to Save Lives, Solstice
07mTransforming Big Data into Real-Time Product Insights