Azure Cosmos DB
Globally distributed, multi-model distributed NoSQL database with turnkey multi-region writes and guaranteed single-digit millisecond latency.
Globally distributed, multi-model distributed NoSQL database with turnkey multi-region writes and guaranteed single-digit millisecond latency.
Coming from Google Cloud?
Cosmos DB matches Firestore for document NoSQL and Spanner for global multi-region active-active writes, while offering multiple APIs (Mongo, Cassandra, Gremlin).
- Fully managed, serverless, horizontally partitioned distributed database.
- Automatic indexing of all document properties by default.
- Event streaming via Change Feed (equivalent to Firestore document triggers / BigQuery streams).
- Consistency: Spanner uses TrueTime hardware clocks for global strong consistency; Cosmos DB offers 5 tunable consistency levels.
- APIs: Firestore is strictly a document database; Cosmos DB supports NoSQL, MongoDB, Cassandra, Gremlin, and Table.
- Billing: Firestore bills per document read/write; Cosmos DB bills per Request Unit (RU/s) throughput.
What is Azure Cosmos DB?
Azure Cosmos DB is Microsoft's globally distributed, multi-model database service. Built on an Atom-Record-Sequence (ARS) core engine, it supports multiple wire-protocol APIs (NoSQL/Core, MongoDB, Cassandra, Gremlin, Table) while offering 5 well-defined consistency levels, turnkey multi-region replication, and contractual SLAs for availability, latency, throughput, and consistency.
Decision Guide: When to Choose This Service
Use Azure Cosmos DB for global web/mobile applications requiring active-active writes worldwide, high-throughput IoT telemetry, real-time personalization, and applications needing 99.999% Five-Nines availability with sub-10ms latency SLAs.
Do NOT use Azure Cosmos DB for complex relational OLTP systems requiring heavy multi-table cross-partition ACID transactions or ad-hoc star-schema analytical warehousing (use Azure SQL or Azure Synapse instead).
Core Architectural Pillars
Read and write locally in any region with sub-10ms P99 latency and 99.999% availability.
Strong, Bounded Staleness, Session (default), Consistent Prefix, and Eventual consistency.
Normalized throughput currency; 1 point read of a 1 KB item equals exactly 1 RU.
Logical partitions (max 20 GB) mapped across physical SSD shards (max 50 GB and 10k RU/s).
Continuous, ordered log of document modifications powering event-driven microservices.
Built-in DiskANN vector indexing for semantic similarity search in generative AI RAG.