1 · When Cosmos DB Fits

• Global distribution — replicate data to 60+ Azure regions with a single click
• Multi-model — document, key-value, graph, column-family, table
• <10 ms latency — SLA-backed P99 reads <10 ms, writes <15 ms
• Elastic scale — millions of RU/s, petabytes of storage
• Schema-agnostic — automatic indexing of all properties by default

Choose Cosmos DB when your workload is globally distributed, high-throughput, and latency-sensitive with flexible schemas.

2 · Available APIs

3 · Partition Keys & Data Modeling

Partition Key Best Practices

• High cardinality — many distinct values (e.g., /userId, /tenantId)
• Even distribution — avoid keys that funnel traffic to one partition (hot partition)
• Query affinity — most queries include the partition key → single-partition reads

Anti-Patterns

• ❌ Low-cardinality keys (/status with 3 values) → hot partitions
• ❌ Timestamp as key → write-heavy on latest partition
• ❌ Frequent cross-partition queries → high RU cost, fan-out

Hierarchical partition keys (preview): combine up to 3 levels (e.g., /tenantId + /userId) for finer distribution.

4 · Consistency Levels

5 · RU/s & Throughput Models

A Request Unit (RU) = cost of reading a 1-KB item by ID. Writes ≈ 5× reads; queries vary by complexity.

RU Planning Tips

• Use the Capacity Calculator or SDK's x-ms-request-charge header
• Index fewer paths → lower write RU cost
• Point reads (ID + partition key) = cheapest operation (1 RU for 1 KB)
• Cross-partition queries cost proportionally to partitions touched

6 · Multi-Region Writes & Conflict Resolution

• Single-region write — one write region, multiple read replicas; automatic failover
• Multi-region write — all regions accept writes; lower write latency globally

Conflict Resolution Policies

• Last Writer Wins (LWW) — default; uses _ts or custom path
• Custom (stored procedure) — app-defined merge logic
• Async — conflicts queued in a feed for manual resolution

Multi-region writes add ~25% cost (additional RU charge per replica write).

7 · Cost Optimization

• TTL (Time to Live) — auto-expire items; no RU cost for deletion
• Analytical Store — column-oriented store for OLAP; no impact on transactional RUs
• Reserved Capacity — 1-yr or 3-yr reservations; up to 65% savings
• Shared throughput — database-level RU/s shared across containers (400 RU minimum)
• Composite indexes — reduce ORDER BY / multi-filter RU cost
• Exclude unused paths from indexing policy

8 · When NOT Cosmos DB

• ❌ Complex relational joins & multi-table transactions → Azure SQL
• ❌ Full-text search as primary need → Azure AI Search
• ❌ Budget-constrained, single-region, simple key-value → Table Storage / Redis
• ❌ Large analytical queries over TBs → Synapse Analytics
• ❌ Need stored procedures in T-SQL → SQL Database

9 · Real-World: Global E-Commerce Product Catalog

10 · Knowledge Check