Computation-Proximate Architecture: An Actor-Based Pattern for Real-Time Calculations Over Billion-Row Datasets

Organizations processing historical datasets with billions of rows face a fundamental architectural choice: move data to computation (the traditional query-centric approach) or move computation to data. This paper presents the \textbf{Computation-Proximate Pattern}, an architecture that combines actor-based in-memory computation with tiered caching and row-key value stores to achieve sub-15ms response times over billion-row datasets. The pattern eliminates unnecessary data movement by resolving 70\% of data lookups from a two-level cache hierarchy---50\% from in-process L1 caches ($<$1$\mu$s) and 20\% from distributed L2 caches ($<$500$\mu$s)---and serving the remaining 30\% through asynchronous parallel key-value lookups ($<$10ms per get). We formalize the pattern's components, analyze its theoretical properties including expected latency under the order statistics of parallel lookups, and validate it empirically on two production systems: a financial calculation engine processing 8 million records daily against 9 billion historical rows, and a telecommunications middleware handling 30,000 concurrent connections over 4 billion historical events. The median end-to-end latency is 3.2ms (p99: 12.1ms). Comparative analysis against batch-processing and direct-query approaches on the same workload demonstrates order-of-magnitude latency reduction at significantly lower infrastructure cost. The pattern is technology-agnostic and applicable to any domain requiring real-time calculations over massive historical datasets.