We present a columnar storage engine that learns its physical design online. Rather than fixing partitioning and sort order at load time, the engine maintains lightweight models of the observed query distribution and reorganizes data only when a cost model predicts net benefit under realistic write amplification. On TPC-H and three production traces, the approach reduces scan latency by 2.1–3.4× over statically tuned baselines while keeping reorganization overhead under 6%.

Learned indexes assume a stationary key distribution — an assumption production workloads routinely violate. We introduce a drift detector that monitors prediction error in situ and triggers targeted model repair instead of full retraining. The method bounds worst-case lookup degradation during shift and recovers index accuracy 9× faster than periodic rebuilds, validated across six evolving real-world datasets.

Compaction scheduling in LSM-tree stores trades write amplification against read performance, usually via hand-tuned heuristics. We derive an analytical cost model that lets the engine speculatively compact ranges ahead of predicted read demand. The model is cheap to evaluate per-flush and improves p99 read latency by up to 28% on write-heavy workloads without raising total write volume.

Self-tuning systems often report median-case wins while hiding tail regressions caused by background reorganization. We propose a benchmarking protocol that isolates and reports adaptation-induced tail latency as a first-class metric, and release an open harness so adaptive-database claims can be reproduced and compared fairly.

Building a secondary index normally pauses ingest or doubles write cost. We show how to construct learned indexes incrementally, amortizing the work across the write path so that worst-case per-operation overhead stays bounded. The engine reaches full index quality with no measurable ingest stall across four production-scale traces.

Range scans dominate analytical workloads yet most learned indexes optimize for point lookups. We fit piecewise-linear models directly to predicate selectivity, letting the optimizer skip blocks with provable bounds. The approach cuts bytes scanned by 35–58% on selective ranges with sub-kilobyte model footprints.

Recursive model indexes are accurate but memory-hungry, which rules them out for embedded and edge deployments. RMI-Lite quantizes and prunes the model hierarchy to fit a fixed memory budget while bounding lookup error, making learned indexing practical on devices with a few megabytes of RAM.

Reproducible evaluation of adaptive indexes needs workloads whose distribution shifts in controlled, documented ways. We release a parameterized generator that emits key streams with tunable drift rate, magnitude, and seasonality, plus six reference traces now used by several follow-on papers.

Learned Bloom filters promise lower false-positive rates by modeling the key set, but their guarantees quietly assume balanced queries. We characterize how negative-query skew erodes those gains and give a simple backup-filter sizing rule that restores worst-case bounds without discarding the learned component.

Column order and encoding decide how well a scan rides the cache hierarchy, yet they are usually picked by rule of thumb. We model the interaction between layout and CPU cache behavior and select per-column encodings that keep hot ranges resident, yielding consistent scan-throughput gains on commodity hardware.

An early-PhD survey mapping where machine-learned components had begun to replace hand-tuned heuristics across indexing, compaction, layout, and cardinality estimation. The taxonomy and open-problem list shaped the direction of the dissertation that followed.