Cassandra vs MongoDB

Quick verdict

Apache Cassandra is an open source distributed database from the Apache Software Foundation, released under the Apache License 2.0. Data is stored in tables with a partition key and clustering columns and queried with CQL (Cassandra Query Language), which looks like SQL but deliberately has no joins. Every node can accept reads and writes; there is no primary. The current release line is 5.0 (5.0.9, released 7 August 2026, per the download page), with 4.1 and 4.0 still maintained. Cassandra 5.0 added Storage Attached Indexes (SAI), a vector data type with similarity functions, trie memtables and SSTables, and the Unified Compaction Strategy.

MongoDB is a document database. Data is stored as BSON documents in collections and queried with the MongoDB Query API. A replica set has one primary that takes writes and secondaries that replicate from it; sharding spreads data across replica sets. MongoDB Community Server is under the Server Side Public License, the current manual release is 9.0, and MongoDB, Inc. runs MongoDB Atlas on AWS, Azure and Google Cloud.

Both are NoSQL databases that scale out, so they are a realistic choice for the same project. They differ most in how you design data, how replication and consistency work, and how much query flexibility you keep after launch. For the broader relational versus non-relational question, see SQL vs NoSQL.

Side by side

Key differences

Data modelling: query-first tables versus documents

The Cassandra documentation is direct about modelling: "joins are not supported in Cassandra so all required fields (columns) must be grouped together in a single table", data "is modeled around specific queries", and duplicating data across tables (denormalisation) is the normal way to serve several access patterns. The partition key decides which nodes hold the data; clustering columns sort rows inside a partition.

MongoDB also encourages modelling for your queries, usually by embedding related data in one document, but you can add a secondary index later and query on new fields without restructuring the collection. That flexibility is the main practical difference between the two after launch.

Recent orders for a customer, modelled both ways:

-- Cassandra (CQL): a table designed for this one query
CREATE TABLE orders_by_customer (
  customer_id uuid,
  order_date  timestamp,
  order_id    uuid,
  total       decimal,
  PRIMARY KEY ((customer_id), order_date, order_id)
) WITH CLUSTERING ORDER BY (order_date DESC, order_id ASC);

SELECT order_id, order_date, total
FROM   orders_by_customer
WHERE  customer_id = 5b6962dd-3f90-4c93-8f61-eabfa4a803e2
LIMIT  10;

// MongoDB (mongosh): one collection plus an index
db.orders.createIndex({ customer_id: 1, order_date: -1 })
db.orders.find({ customer_id: 42 }, { order_id: 1, order_date: 1, total: 1 })
         .sort({ order_date: -1 })
         .limit(10)

If you later need "orders by product", Cassandra's guidance points you to a second table (for example orders_by_product) written alongside the first, or to an index; in MongoDB you would typically add an index on the existing collection.

Replication and tunable consistency

Cassandra replication is leaderless. The documentation states that "write operations are always sent to all replicas, regardless of consistency level", and the consistency level only decides how many replicas must acknowledge before the client gets a reply. Levels include ONE, TWO, THREE, QUORUM, ALL, LOCAL_QUORUM, EACH_QUORUM, LOCAL_ONE and the write-only ANY. Reads see the latest write when W + R > RF (write level plus read level greater than the replication factor). NetworkTopologyStrategy sets a replication factor per data centre, and hinted handoff, read repair and anti-entropy repair bring replicas back in line.

-- Cassandra (cqlsh): keyspace replicated to two data centres
CREATE KEYSPACE shop WITH replication =
  {'class': 'NetworkTopologyStrategy', 'dc1': 3, 'dc2': 3};
CONSISTENCY LOCAL_QUORUM;

MongoDB uses a primary per replica set. Writes go to the primary, secondaries replicate from it, and an eligible secondary calls an election if the primary is unreachable. Consistency is controlled with write concern (default w: "majority" on most replica sets) and read concern. The practical difference: Cassandra can keep accepting writes in every data centre during a partition if you choose a local consistency level, while MongoDB writes for a shard wait for its primary.

Write path

Cassandra uses a log-structured merge design. Per the storage engine documentation, each write is appended to a commit log on disk and stored in an in-memory memtable; when the memtable reaches its limit it is flushed to an immutable SSTable, and compaction later merges SSTables and processes updates and deletes. The documentation notes the trade-off: appends are cheap, but compaction adds write amplification, and a partition is typically spread across several SSTables until compaction merges them.

An INSERT in Cassandra is also an upsert: the CQL documentation says that, unlike SQL, INSERT does not check whether the row exists and updates it if it does. Deletes write tombstones that are removed at compaction, which is one reason data models avoid heavy delete patterns.

MongoDB writes through its storage engine with journaling, and documents are updated in place from the application's point of view, so read-modify-write patterns on one document are simple and single-document writes are atomic.

Transactions

Cassandra has no general multi-row transactions. It offers lightweight transactions through IF NOT EXISTS and IF conditions, which use Paxos and which the documentation warns carry "a non-negligible performance cost" and should be used sparingly. Logged BATCH statements ensure all mutations eventually complete or none do, but isolation applies only within a single partition.

-- Cassandra (CQL): conditional insert using a lightweight transaction
INSERT INTO users (username, email) VALUES ('ada', 'ada@example.com')
IF NOT EXISTS;

MongoDB supports multi-document ACID transactions on replica sets and sharded clusters with commit and abort. MongoDB itself advises designing documents so that most operations do not need them, but they are available when an operation spans documents, for example moving stock between two warehouses.

Managed services: Astra DB, Amazon Keyspaces and Atlas

Astra DB is DataStax's managed service built on Cassandra, offering CQL and a JSON Data API with vector search. IBM announced its acquisition of DataStax in February 2025; datastax.com now redirects to IBM's product pages and the Astra DB documentation carries an IBM copyright, so check current terms with IBM.

Amazon Keyspaces (for Apache Cassandra) is a serverless, Cassandra-compatible AWS service, not Apache Cassandra itself. AWS documents compatibility with the CQL 3.11 API and lists features it does not support, including CREATE INDEX, materialised views, user-defined functions and aggregates, triggers and TRUNCATE; compaction, compression and similar settings are managed by AWS.

MongoDB Atlas is run by MongoDB, Inc. on AWS, Azure and Google Cloud, with Free, Flex and Dedicated tiers. Because Atlas is the vendor's own service, new MongoDB features generally reach it directly; with Cassandra you choose between self-managing the Apache project or a provider's distribution or compatible service.

Pricing and licensing

Licences. Apache Cassandra is free under the Apache License 2.0, a permissive OSI-approved licence. MongoDB Community Server is free under the SSPL v1.0, which is not OSI-approved; commercial licences come with MongoDB Enterprise Advanced through sales. We describe licence terms only and do not give legal advice.

MongoDB Atlas. Listed on the MongoDB pricing page in October 2026, in USD: Free tier with 512 MB of storage; Flex at 0.011 USD per hour, capped at 30 USD per month; Dedicated clusters from 0.08 USD per hour (M10).

Amazon Keyspaces. The AWS pricing page describes on-demand mode (pay per read and write request unit) and provisioned mode (capacity per second, with auto scaling), plus storage per GB-month. A read unit covers up to 4 KB at LOCAL_QUORUM and a write unit up to 1 KB. New customers get a three-month free tier of 30 million on-demand write units, 30 million read units and 1 GB of storage per month. Rates vary by Region, so use the AWS calculator for estimates.

Astra DB. IBM's pages offer a free trial; we could not find a current public price list, so check with IBM.

Pricing checked on the vendors' official pages on 7 October 2026. Prices change; confirm before buying.

Where each one leads

Limitations

When to choose each

When neither is right

• Relational data with joins, constraints and SQL reporting: use a relational database; see MongoDB vs PostgreSQL or MongoDB vs MySQL.
• A serverless key-value store with pay-per-request pricing entirely inside AWS: see DynamoDB vs MongoDB.
• Sub-millisecond caching, sessions or counters: an in-memory store fits better; see Redis vs MongoDB.
• Large analytical scans and aggregations: a column-oriented analytics engine or data warehouse is designed for this rather than an operational database.

Final recommendation

Frequently asked questions

Sources

• Apache Cassandra downloads
• Cassandra 5.0: new features
• Cassandra documentation: Data modelling introduction
• Cassandra documentation: Dynamo (replication and consistency)
• Cassandra documentation: Storage engine
• Cassandra documentation: Data manipulation (CQL)
• Amazon Keyspaces: What is Amazon Keyspaces
• Amazon Keyspaces: Supported Cassandra APIs
• Amazon Keyspaces pricing
• Astra DB Serverless documentation
• IBM newsroom: IBM to acquire DataStax
• MongoDB manual: Write concern
• MongoDB manual: Transactions
• MongoDB Community Edition licensing
• MongoDB Atlas pricing

Checked October 2026.

How we research comparisons: our editorial method.