NoSQL Types and Comparison

Not Everything Fits in a Table

A social network stores millions of users, and every user has a different combination of profile fields, privacy settings, and connected friends. A product catalog has phones with specs like "battery capacity" and shoes with specs like "heel height" completely different attributes. A gaming leaderboard needs to update 50,000 scores per second.

Try to model all three in a relational schema. The phone shoe problem means either dozens of NULL columns or a complex EAV (Entity Attribute Value) anti pattern. The social graph means recursive self joins that bring PostgreSQL to its knees. The leaderboard means locking and write contention that kills response times.

Not all data fits in tables. NoSQL databases were built for the shapes that relational models handle poorly.

Why NoSQL Emerged: RDBMS Limitations at Scale

Document Stores MongoDB

Model: Collections of JSON like documents. Each document can have its own schema.

// products collection
{
  "_id": "prod_001",
  "name": "iPhone 15 Pro",
  "category": "phone",
  "specs": {
    "battery_mah": 3274,
    "chip": "A17 Pro",
    "storage_options": [128, 256, 512, 1024]
  },
  "price": 999.00,
  "tags": ["apple", "smartphone", "5G"]
}

{
  "_id": "prod_002",
  "name": "Air Max 90",
  "category": "shoe",
  "specs": {
    "heel_height_mm": 32,
    "material": "leather/mesh",
    "sizes": [7, 8, 9, 10, 11, 12]
  },
  "price": 110.00
}

Query example:

// Find all phones under $800 with storage over 256GB
db.products.find({
  category: "phone",
  price: { $lt: 800 },
  "specs.storage_options": { $gt: 256 }
})

Best use cases: Product catalogs, content management systems (CMS), user profiles, event logging, mobile app backends.

Key Value Stores Redis

Model: A giant hash map. Every value is stored and retrieved by a unique key. Values can be strings, lists, sets, hashes, sorted sets.

SET session:user_4821  '{"user_id":4821,"role":"admin","expires":1720000000}'
GET session:user_4821
→ '{"user_id":4821,"role":"admin","expires":1720000000}'

SETEX rate_limit:ip_192.168.1.1  60  "42"     -- expires in 60 seconds
INCR  rate_limit:ip_192.168.1.1               -- atomic increment

ZADD  leaderboard  98500 "player_alice"
ZADD  leaderboard  87200 "player_bob"
ZREVRANGE leaderboard 0 9 WITHSCORES          -- Top 10 players
→ 1) "player_alice"  2) "98500"
   3) "player_bob"   4) "87200"

Speed: Redis operates entirely in RAM typical operations complete in under 1 millisecond.

Best use cases: Session storage, caching (cache aside pattern), rate limiting, real time leaderboards, pub/sub messaging queues.

Column Stores Apache Cassandra

Model: Data is organized by rows with dynamic columns. Rows are grouped into partitions identified by a partition key, optimized for fast writes and time series reads.

-- Cassandra CQL (similar to SQL)
CREATE TABLE sensor_readings (
    device_id    UUID,
    recorded_at  TIMESTAMP,
    temperature  FLOAT,
    humidity     FLOAT,
    PRIMARY KEY (device_id, recorded_at)
) WITH CLUSTERING ORDER BY (recorded_at DESC);

-- Write millions of readings per second
INSERT INTO sensor_readings (device_id, recorded_at, temperature, humidity)
VALUES (uuid(), toTimestamp(now()), 22.4, 61.2);

-- Fast range read for one device
SELECT * FROM sensor_readings
WHERE device_id = 550e8400-e29b-41d4-a716-446655440000
  AND recorded_at >= '2024-01-01'
  AND recorded_at <  '2024-02-01';

Cassandra distributes data across a ring of nodes. There is no single point of failure. Writes go to multiple nodes simultaneously for fault tolerance.

Best use cases: IoT sensor data, time series metrics, audit logs, messaging systems, analytics at massive scale (Netflix, Apple, Instagram use Cassandra).

Graph Databases Neo4j

Model: Data is stored as nodes (entities) and edges (relationships), each with properties. Relationships are first class citizens, not foreign keys.

Nodes:  (Alice:Person), (Bob:Person), (Python:Language), (DataCo:Company)
Edges:  Alice -[FRIENDS_WITH]-> Bob
        Alice -[KNOWS]-> Python
        Bob   -[WORKS_AT]-> DataCo
        DataCo-[USES]-> Python

Cypher Query Find Alice's friends who work at companies using Python:

MATCH (alice:Person {name: "Alice"})
      -[:FRIENDS_WITH]->(friend:Person)
      -[:WORKS_AT]->(company:Company)
      -[:USES]->(lang:Language {name: "Python"})
RETURN friend.name, company.name

-- Output:
-- friend.name  | company.name
-- -------------|-------------
-- Bob          | DataCo

The same query in SQL requires 4 self joins with recursive traversal orders of magnitude slower for deep graph queries.

Best use cases: Social networks, recommendation engines, fraud detection, knowledge graphs, network topology, identity and access management.

CAP Theorem

Eric Brewer's CAP Theorem states that a distributed system can guarantee at most two of these three properties simultaneously:

           Consistency (C)
               /\
              /  \
             /    \
            /      \
           /        \
Availability(A) ---- Partition
                     Tolerance (P)

CA: Traditional RDBMS (single node)    — Not partition tolerant
CP: MongoDB, HBase, Zookeeper          — Sacrifices availability during partitions
AP: Cassandra, CouchDB, DynamoDB       — Sacrifices consistency (eventual)

In a distributed system, network partitions will happen. You must choose between Consistency and Availability when they do.

BASE vs ACID

Most NoSQL databases follow BASE semantics instead of ACID:

Eventual consistency means: if no new writes are made, all replicas will eventually converge to the same value. You might read stale data immediately after a write but you will read the correct data within milliseconds to seconds.

For a social media "like" count, this is perfectly acceptable. For a bank balance, it is not.

Master Comparison Table

Key Takeaways

• NoSQL does not mean "no SQL" it means "Not Only SQL." Many NoSQL databases have query languages. It means non relational.
• Document stores (MongoDB) shine for flexible schemas and nested data.
• Key value stores (Redis) are unmatched for speed use them for caching and sessions.
• Column stores (Cassandra) handle massive write throughput and time series data.
• Graph databases (Neo4j) are purpose built for relationships social networks, recommendations, fraud detection.
• CAP Theorem forces a tradeoff in distributed systems understand whether your use case needs consistency or availability during a network partition.
• BASE (eventual consistency) is acceptable for many use cases but never for financial transactions.

Choosing a database is choosing a tradeoff. The best database is the one whose tradeoffs align with your application's actual requirements not the most popular one.