PROVIDER Β· Linearly scalable wide-column distributed database
Apache Cassandra
Apache Cassandra blends Dynamo distribution with the Bigtable columnar model into a wide-column database whose peer-to-peer architecture, linear write scaling, and multi-datacenter replication fit massive ingestion workloads.
Pros
Cons
Product Information
Detailed Notes
Overview
Apache Cassandra was open sourced by Facebook in 2008, blending the decentralised distribution of Amazon Dynamo with the columnar model of Google Bigtable, becoming an Apache top-level project in 2010 and used by Netflix, Apple, and Discord for very high write throughput.
For Backend & Database work the core is a peer-to-peer design with linear scaling, since any node accepts reads and writes, adding nodes raises throughput, multi-datacenter replication gives natural cross-region resilience, and write performance grows with cluster size, ideal for logs, IoT, messaging, and telemetry.
Accept the trade-offs, because only tunable eventual consistency is available, joins and cross-partition transactions are absent, data models must be built per query denormalised, and operations require tuning compaction, repair, and GC, so relational usage becomes painful.
Core Strengths
- Peer-to-peer architecture lets any node read and write with no single point of failure
- Write throughput scales linearly with nodes into hundreds of thousands per second
- Native multi-datacenter replication provides cross-region resilience without middleware
- Tunable consistency levels trade latency against guarantees per query
Product Ecosystem
Data model and CQL
SQL-like CQL syntax with tables designed per query, where the partition key governs distribution and hotspots and denormalisation is normal.
Cluster and replication
Partition-key hashing distributes data while replication factor and consistency level set durability and latency, requiring topology planning across datacenters.
Operations
Compaction strategy, repair cadence, GC, and tombstone cleanup determine stability, with DataStax and community tools for monitoring and repair.
Drawbacks
- No joins or cross-partition transactions, pushing complex queries into the app or extra components
- Models must be designed per query, so requirement changes often mean new tables with dual writes
- Operations are complex, with poor repair and compaction settings collapsing performance
- Eventually consistent only, so strong consistency needs workarounds or another engine
Use Cases
- Log, telemetry, and IoT ingestion (β β β β β ) - Linear write scaling
- Multi-region messaging and records (β β β β β ) - Native datacenter replication
- Per-user profile and recommendation data (β β β β ) - Wide rows fit naturally
- Complex joins and reporting (β ) - No joins, needs a separate pipeline
Pricing
| Plan | Price | Scope |
|---|---|---|
| Apache Cassandra | $0 | Open source, self-managed |
| Managed Cassandra | Per node or usage | Operations-free with backup and monitoring |
| Cloud provider managed | Per request unit and storage | CQL compatible with trimmed features |
The open source edition is free while self-hosting hides real cost, needing three nodes minimum and dedicated tuning of repair and compaction, so managed pricing often wins below five nodes or without in-house expertise.
FAQ
-
Cassandra or MongoDB? Cassandra for write throughput and multi-datacenter resilience, MongoDB for flexible queries, aggregation, and faster iteration, both without joins but Cassandra leaning write-side.
-
Can it be the primary database? Yes with eventual consistency and query-driven modelling accepted, while money and inventory transactions needing strong consistency belong in a relational engine.
-
What is the minimum cluster size? Production wants at least three nodes with replication factor three and additional sites for cross-region, as one or two nodes cannot survive a failure.