Application of B-Trees in Real-Time Transaction Processing

Abstract

Real-time transaction processing (RTTP) systems must complete database operations within strict latency budgets while sustaining high throughput, even under bursty, skewed workloads. Index structures dominate the critical path of these systems because almost every transactional read, write, or predicate check touches an index. This manuscript examines the practical application of B-Trees and close variants to RTTP, focusing on predictability, concurrency, and deadline-aware behavior. We begin by revisiting B-Tree fundamentals—node fan-out, height bounds, and page layout—and explain why those properties make B-Trees attractive for bounded-time access. We then review techniques that convert theoretical advantages into end-to-end deadline reliability: latch coupling, crabbing, B-link side pointers, optimistic and lock-free traversals, and append-friendly logging.

Building on these ingredients, we propose a deadline-aware B-Tree (DABT) design that (1) pins upper levels in memory, (2) pre-splits hot pages during slack time, (3) uses deadline-aware latch acquisition with time budgets, and (4) aligns group-commit windows to transaction deadline bins. A discrete-event simulator and OLTP-style microbenchmarks (read-heavy and write-intensive mixes under Zipfian key skew) are used to evaluate baseline latch-coupled B-Trees, B-link trees, and DABT. Results show that DABT reduces p99 index latency by 29–44% relative to baselines, cuts deadline-miss rates by more than half, and improves overall throughput by ~11–20% without sacrificing serializability. We analyze statistical significance with repeated trials and report mean±SD across key metrics. The findings indicate that carefully engineered B-Tree variants remain a compelling choice for predictable, real-time OLTP on modern NVMe/DRAM systems.