Polyglot Persistence

Most real systems use multiple databases, each chosen for a specific access pattern.

Common Access Patterns

Pattern	Description	Example	Best Fit
Point reads/writes	Get/set a single record by key	Fetch user by ID	KV store, relational
CRUD	Create, Read, Update, Delete	Update a tweet	Relational (Postgres)
Aggregations	Compute over millions of rows	AVG price by region	Columnar (Redshift)
Graph traversals	Multi-hop relationship walks	Friends-of-friends	Graph DB (Neo4j)
Full-text search	Keyword/relevance search	"Search tweets about X"	Elasticsearch
Time-series	Append-heavy, ordered by time	Metrics, IoT, logs	TimescaleDB, InfluxDB
Wide-column / sparse	Many columns, mostly null	User preferences	Cassandra, HBase

---

Example: Twitter's Database Architecture

graph TD
    App["Twitter App"] -->|"CRUD"| PG["PostgreSQL<br/>Users, Tweets, Likes"]
    App -->|"recommendations"| Graph["Neo4j / Custom Graph<br/>Who To Follow"]
    App -->|"search"| ES["Elasticsearch<br/>Tweet Search"]
    App -->|"cache"| Redis["Redis<br/>Timeline Cache"]
    PG -->|"CDC / ETL"| RS["Redshift<br/>Analytics Dashboard"]

    style PG fill:#1a3a4a,stroke:#44aaff
    style Graph fill:#3a1a4a,stroke:#aa44ff
    style ES fill:#4a3a1a,stroke:#ffaa44
    style Redis fill:#1a4a1a,stroke:#44ff44
    style RS fill:#4a1a2a,stroke:#ff44aa

---

Keeping Data in Sync

With multiple databases, consistency between them is a hard problem:

Strategy	How It Works	Tradeoff
Dual writes	App writes to both DBs	Fragile — if one write fails, data drifts
CDC (Change Data Capture)	One DB is source of truth, changes stream to others via Debezium/Kafka	Eventually consistent, but reliable
Event sourcing	All changes go to event log first, each DB builds its own view	Most robust, but complex to implement

CDC is the most common pattern in practice. The primary relational DB is the source of truth; everything else is a derived view.

graph LR
    PG["PostgreSQL<br/>(Source of Truth)"] -->|"CDC stream<br/>(Debezium + Kafka)"| Neo["Neo4j"]
    PG -->|"CDC stream"| ES["Elasticsearch"]
    PG -->|"CDC stream"| RS["Redshift"]
    PG -->|"Cache invalidation"| Redis["Redis"]

    style PG fill:#1a3a4a,stroke:#44aaff

"One DB to rule them all?" — Systems like CockroachDB, SurrealDB, and even PostgreSQL (with extensions) try to handle multiple patterns. But at scale, specialized systems with purpose-built storage engines will always outperform a generalist. The industry consensus: right tool for each job + CDC to keep them in sync.