Modern data processing systems—databases, analytics engines, vector stores, and stream processors—hide an extraordinary amount of performance engineering beneath their abstractions. In this talk, we “open the lid” on data systems and reveal the low‑level performance tricks they rely on to achieve staggering throughput: vectorized execution, branch‑aware operators, cache‑optimized data layouts, late materialization, compressed‑buffer processing, adaptive execution paths, and microarchitectural tuning. We’ll break down what really happens between a high‑level query and the CPU pipeline—and why systems spend so much energy manipulating data before doing any useful computation.

Drawing on data processing systems research, we explore the surprising bottlenecks hidden inside operators, memory hierarchies, and parallelization strategies. We abstract these insights into concrete techniques you can apply directly in your own codebases: when to favor columnar layouts, how to reduce branch mispredictions, when compression accelerates computation, how to leverage vectorization safely, and how modern profiling tools can expose CPU‑level inefficiencies that traditional metrics miss.

Whether you build databases, ML pipelines, backend services, or high‑performance libraries, this talk will give you a new mental model of how data really flows through modern hardware—and how to “liberate" the best ideas from the fastest systems in the world to make your code faster.