Databases — Roadmap

Every AI system needs somewhere to put data — structured, vector, graph, and NoSQL.

1. Relational (MySQL, Postgres)

• Relational model, normalization
• SQL fundamentals: joins, aggregations, window functions
• Indexing and query planning
• Transactions and ACID guarantees
• When relational is (and isn't) the right tool for ML metadata/feature storage

2. Vector Databases

• What a vector database actually stores and searches (embeddings + approximate nearest neighbor)
• Distance metrics: cosine, dot product, Euclidean
• Indexing algorithms: HNSW, IVF, product quantization
• ChromaDB — local/embedded vector store for prototyping
• Production-scale options (Pinecone, Weaviate, Milvus, pgvector) and how to choose
• Hybrid search: combining vector similarity with keyword/metadata filters

3. Graph Databases

• Property graph model: nodes, edges, properties
• Neo4j and Cypher query language basics
• When a graph model beats relational (highly connected data, multi-hop queries)
• GraphRAG — using a knowledge graph to ground LLM retrieval

4. NoSQL Databases

• Why NoSQL exists: relaxing the relational model's schema/joins/transactions for scale and flexibility
• Document stores (MongoDB) — nested, variable-shape records
• Key-value stores (Redis, DynamoDB) — fastest possible lookup by key, the standard caching layer
• Wide-column stores (Cassandra) — massive, distributed write throughput
• When to reach for NoSQL over relational, and the real tradeoff either way

Common Problems & SOTA Solutions

• Vector search doesn't scale past X million vectors → approximate indexing (HNSW/IVF), sharding
• Pure vector search misses exact keyword matches → hybrid search + re-ranking
• Stale embeddings after content updates → versioned re-indexing pipelines (see RAG — Common Failure Modes)
• Multi-hop reasoning fails with flat vector retrieval → GraphRAG / knowledge graph augmentation