Database Paradigms · Shreedhar Kodate

The database landscape has fragmented dramatically — no single paradigm fits all workloads. Understanding the data model, query capabilities, and operational characteristics of each paradigm is essential for matching storage technology to access patterns.

Database Taxonomy

graph TD DB[Database Systems] DB --> Relational[Relational - SQL\nPostgres, MySQL, Oracle, CockroachDB] DB --> Document[Document\nMongoDB, Firestore, CouchDB] DB --> KeyValue[Key-Value\nRedis, DynamoDB, etcd] DB --> WideColumn[Wide-Column\nCassandra, HBase, Bigtable] DB --> Graph[Graph\nNeo4j, Amazon Neptune, TigerGraph] DB --> TimeSeries[Time-Series\nInfluxDB, TimescaleDB, Prometheus] DB --> Vector[Vector\nPinecone, Weaviate, pgvector, Qdrant] DB --> Search[Search Engine\nElasticsearch, OpenSearch, Typesense] DB --> Columnar[Columnar Analytical\nClickHouse, Redshift, BigQuery, DuckDB] style Relational fill:#dbeafe,stroke:#2563eb style Vector fill:#fef3c7,stroke:#d97706

Relational Model

graph TD subgraph Schema[Relational Schema - Orders System] Users[Users\nuser_id PK\nname\nemail\ncreated_at] Orders[Orders\norder_id PK\nuser_id FK\nstatus\ntotal_amount\ncreated_at] OrderItems[Order Items\nitem_id PK\norder_id FK\nproduct_id FK\nquantity\nunit_price] Products[Products\nproduct_id PK\nname\nsku\nprice] Users -->|1 to many| Orders Orders -->|1 to many| OrderItems OrderItems -->|many to 1| Products end subgraph SQL_Strengths[SQL Strengths] JOINS[Complex JOINS across tables] ACID2[Full ACID transactions] Norm[Normalized - no duplication] Flex[Ad-hoc query flexibility] end

NoSQL Data Models

graph LR subgraph DocumentModel[Document Model - MongoDB] Doc[Order Document\nbrace user_id name email\n orders array\n item product_id qty price\nclose brace\nDenormalized - all in one doc] style Doc fill:#dcfce7,stroke:#16a34a end subgraph WideColumnModel[Wide-Column - Cassandra] Row[Row Key: user_id dash order_id\nColumn Families:\n order_info: status amount\n items: product quantity\nDesigned for specific query patterns] style Row fill:#dbeafe,stroke:#2563eb end subgraph KVModel[Key-Value - Redis] KV[Key: session:abc123\nValue: JSON blob or binary\nO1 access by exact key\nNo query flexibility] style KV fill:#fef3c7,stroke:#d97706 end

Vector Database Architecture

graph TD subgraph Ingestion[Data Ingestion] Docs[Raw Documents\nImages, Text, Code] Embedder[Embedding Model\nOpenAI, Sentence-BERT, CLIP] Docs --> Embedder Embedder -->|dense vectors| VectorStore[Vector Index\nHNSW - ANN index] end subgraph Query[Similarity Search] Query[Query: user text] QEmbed[Embed query] --> SimilaritySearch[ANN Search\nTop-K nearest vectors] VectorStore --> SimilaritySearch SimilaritySearch --> Results[Top-K similar documents] end Query --> QEmbed style VectorStore fill:#fef3c7,stroke:#d97706,stroke-width:2px style SimilaritySearch fill:#dcfce7,stroke:#16a34a

Key Concepts

Relational (SQL): Organizes data into tables with predefined schemas, enforcing referential integrity via foreign keys. ACID transactions span multiple tables via JOINs. SQL is the most powerful query language — ad-hoc queries, aggregations, and complex joins are native. Vertical scaling bottleneck for write-heavy workloads beyond what a single node can handle.
Document Database: Stores data as semi-structured documents (JSON/BSON). Documents can contain nested objects and arrays, enabling denormalized data models where related data lives together. Excellent for read-heavy workloads where a single document satisfies a query. Poor for operations that span multiple documents (no joins).
Key-Value Store: The simplest model — a dictionary with O(1) point lookups by key. No schema, no query language beyond get/set/delete. Redis adds data structures (sorted sets, lists, hashes) and expiration. Used for caching, session storage, feature flags, and rate limiting.
Wide-Column (Column-Family): Stores data in rows with dynamic column families. Rows are partitioned and sorted by a composite key. Optimized for high-write throughput and queries on time-series or event data where the query pattern is known upfront. Cassandra's data model forces schema design around queries, not around domain entities.
Graph Database: Stores entities (nodes) and relationships (edges) as first-class citizens with properties. Graph traversal (finding paths, neighbors, subgraphs) is native and performant. Relational databases can model graph data but become expensive as traversal depth increases. Used for social networks, fraud detection, recommendation engines.
Time-Series Database: Optimized for append-only writes of timestamped data points. Uses specialized compression (delta encoding, gorilla compression) and retention policies. Supports time-range aggregations efficiently. Used for metrics, IoT, financial tick data, application telemetry.
Vector Database: Stores high-dimensional dense vectors (embeddings) and supports approximate nearest neighbor (ANN) search. Used for semantic similarity search, recommendation systems, and RAG (Retrieval-Augmented Generation) architectures. HNSW (Hierarchical Navigable Small World) is the dominant index algorithm.

Trade-offs

Paradigm	Query Flexibility	Write Throughput	Schema	Consistency	Best For
Relational	Highest	Moderate	Strict	ACID	Complex queries, transactions
Document	Moderate	High	Flexible	Varies	Hierarchical data, APIs
Key-Value	Point lookup only	Very high	None	Varies	Caching, sessions
Wide-Column	Limited (query design)	Very high	Semi-structured	Tunable	Time-series, write-heavy
Graph	Graph traversal	Moderate	Flexible	ACID	Relationships, networks
Time-Series	Time-range queries	Very high	Structured	Varies	Metrics, telemetry
Vector	Similarity search	Moderate	Fixed-dimension	Eventual	AI search, recommendations

When to Use

Relational: Default for most application data with complex relationships and transaction requirements
Document: Content management, product catalogs, user profiles, API backends with hierarchical data
Key-Value: Caching, session management, rate limiting counters, feature flags
Wide-Column: IoT telemetry, analytics event tables, write-heavy time-series at massive scale
Graph: Social networks, fraud detection, recommendation engines, knowledge graphs
Time-Series: Infrastructure metrics, application monitoring, financial tick data
Vector: Semantic search, RAG, image similarity, recommendation systems using embeddings