NoSQL, CAP and BASE

Different trade-offs for different needs

NoSQL ("not only SQL") databases relax parts of the relational model to gain horizontal scale, flexible schemas or specialised performance. Families: key-value (Redis, DynamoDB), document (MongoDB, Couchbase), wide-column (Cassandra, HBase, Bigtable) and graph (Neo4j). Many favour denormalised data shaped around queries and scale out by partitioning data across nodes. The CAP theorem states that when a network partition occurs, a distributed data store must choose between consistency (every read sees the latest write or returns an error) and availability (every request gets a non-error response). Systems are often described as CP (prefer consistency, such as HBase or etcd) or AP (prefer availability, such as Cassandra in its default settings), though many are tunable. BASE (Basically Available, Soft state, Eventual consistency) describes the AP style, in contrast to ACID. Today the lines blur: many NoSQL systems offer transactions, and distributed SQL databases ("NewSQL", such as Google Spanner, CockroachDB, YugabyteDB) provide ACID transactions at scale.

Picking two during a partition

When the network splits, a distributed database chooses between consistency and availability.

Matching workloads to database families

A rough guide; many systems span several categories.

workload                                         typical choice
-----------------------------------------------  ------------------------------
business transactions, reporting, joins          relational (PostgreSQL, MySQL)
session store, cache, counters, leaderboards     key-value (Redis)
product catalogue with varied nested attributes  document (MongoDB) or JSON columns
IoT/time-series writes at huge scale             wide-column (Cassandra) / time-series DB
friend-of-friend, fraud rings, recommendations   graph (Neo4j)
global ACID transactions at scale                distributed SQL (Spanner, CockroachDB)

NoSQL is not schema-less

Data always has a schema; NoSQL just moves it from the database into application code. Without discipline, documents drift into many shapes that every reader must handle.

त्वरित जाँच: According to the CAP theorem, what must a distributed database choose between during a network partition?

• Speed and cost
• SQL and NoSQL
• Consistency and availability
• Indexes and joins