Choosing partition keys, consistency levels, integrated vector search.

Cosmos DB — Partition Keys, RUs, and Multi-Region

Azure Cosmos DB is a globally distributed NoSQL database with five consistency levels (Strong to Eventual). Request Units (RUs) are the currency — 1 RU ≈ 1 point read of a 1KB document. Choosing the right partition key is critical: high cardinality, even distribution, and alignment with query patterns. Cross-partition queries fan out to all partitions — expensive. Cosmos DB for NoSQL now supports integrated vector search with DiskANN indexing.

Interactive Pipeline Flow

UnderstandImplementOptimise

🗂️ Partition — choose partition key

Partition key determines physical partitioning. Ideal: high cardinality, even write distribution, frequently used in queries as equality filter. Avoid hot partitions (all writes to one value).

Architecture Diagrams

Five Cosmos DB consistency levels and their trade-offs.

Loading diagram...

Key Design Decisions

🗂️Partition key is the most important design decision

A bad partition key cannot be changed without migrating all data. Evaluate: cardinality, write distribution, and query alignment BEFORE creating a container.

💰Point reads are 10-100x cheaper than queries

ReadItemAsync(id, partitionKey) = 1 RU. A cross-partition query with no index can cost 100s of RUs. Design your data so common reads are point reads.

📈Use autoscale for unpredictable workloads

Autoscale charges for actual usage between 10-100% of max RU. For a 1000 RU max with 100 RU actual, you pay for 100 RU. Manual provisioning would cost 1000 RU continuously.

Discussion