Relational or Distributed? A Scaling Study of Frequent Itemset Mining over Urban License Plate Detections for Smart City Platforms

Counting how many vehicles crossed an intersection produces a report; finding which vehicles move together, and through which cameras, reveals the structure of urban traffic. License plate recognition cameras make that structure observable. Sets of cameras frequently shared by the same vehicles expose the corridors that traffic actually uses, and groups of vehicles repeatedly seen together at the same cameras support automated convoy detection for traffic enforcement. Both are frequent itemset mining problems, and they are precisely where relational databases lose ground as data grow. We quantify that loss for the corridor formulation, comparing a self join in indexed PostgreSQL with distributed FP-Growth on a four-worker Apache Spark cluster over 100 million simulated detections from 500 cameras on the road network of Pereira, Colombia, at eight nested scales from 50 thousand to 15 million rows. For reports, histograms, lookups and alerts, PostgreSQL wins at every scale by up to five orders of magnitude, and Spark never leaves a fixed overhead of two to four minutes. Structure mining behaves differently: the relational advantage collapses from 328 to 1.6 times, because the self join grows super-linearly while FP-Growth grows almost linearly. Extrapolated growth exponents place the crossover between 30 and 73 million detections, four to nine days of operation, which tells a city when structural analysis must move off the relational engine.