Learned Indexes: a New Idea for Efficient Data Access #bbuzz
summary
Proposes replacing B‑trees with trained models and shows concrete designs: a neural-network index trained on 200 million logs performs poorly in TensorFlow (incurs two orders of magnitude slower lookups) but a layered recursive model index (cheap local models under a heavier top model) yields much smaller error bounds, constant-time offset predictions, big space savings for hash maps, and halved bloom‑filter size on some integer datasets; requires mostly static data and periodic retraining for inserts.
AI-written summary · about
chapters (6)
- 0:00Introduction to learned indexes
- 3:07Traditional index structures
- 11:32Machine learning basics
- 15:56Applying ML to database indexing
- 24:23Advanced index replacement
- 30:31Benchmarks and Q&A
related talks
30mLearned Indexes in PostgreSQL: Theory & Implementation
50mExploring the benefits of learned indexes in PostgreSQL: Theory and Implementation
41mFrom Embeddings To Index: A Practitioner's Guide To Domain-Adapted Neural Search
27mDropping Into B-Trees
37mRules of Indexing in PostgreSQL
53mAdvanced Indexing Techniques in PostgreSQL
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
23mErik Bernhardsson, CTO, Better Mortgage, NYC 2017