Performance tuning for Azure Cosmos DB - Hasan Savran - NDC Oslo 2024

Physical partitions in Cosmos DB have limits of 50GB storage and 10,000 RUs per server. Understanding these limits is crucial for scaling

Indexing modes and strategies significantly impact costs:

• Everything is indexed by default
• Use composite indexes to optimize queries and save RUs
• Exclude unnecessary properties from indexing
• Index metrics help identify optimization opportunities

SDK connection modes affect performance:

• Direct mode (recommended) uses TCP/TLS
• Gateway mode uses HTTPS/DNS
• Use 64-bit applications and singleton pattern for document clients

Request Units (RUs) optimization:

• Point reads are cheapest way to read data
• Include partition key in queries when possible
• Use partial document updates instead of full updates
• Caching can save RUs for frequently accessed data

Performance best practices:

• Use bulk operations for large data loads
• Implement proper partition key strategy
• Leverage change feed for data synchronization
• Use stream API when possible to reduce memory usage
• Consider serverless for dev/QA environments

Data consistency considerations:

• Avoid “lazy” consistency level
• Session or eventual consistency can be more cost-effective
• Consistency level affects RU consumption

Cost optimization strategies:

• Reserved capacity offers 20-30% savings
• Autoscale for variable workloads
• Monitor and optimize indexing policies
• Use caching for frequently accessed data
• Implement composite indexes strategically