The four major options for a globally distributed database — Spanner (GCP), DynamoDB (AWS), Cosmos DB (Azure), and Aurora DSQL (AWS, GA in 2024) — compared head to head. As the line between NewSQL and NoSQL keeps blurring, this guide lays out how to pick the right one for each workload.
| Item | Spanner | DynamoDB | Cosmos DB | Aurora DSQL |
|---|---|---|---|---|
| Model | NewSQL (SQL) | NoSQL KV/Doc | Multi-model | NewSQL (PG-compatible) |
| SQL | GoogleSQL / PostgreSQL | PartiQL | SQL API + 4 others | PostgreSQL |
| Consistency | External Consistency | Eventual / Strong | 5 levels (Strong / BS / Session / CP / Eventual) | Serializable |
| ACID | Excellent (global) | Excellent (per table) | Excellent (per container) | Excellent (per region) |
| Multi-region | Excellent (5 continents) | Global Tables | Excellent (60+ regions) | 3 regions |
| SLA | 99.999% | 99.999% (Global Tables) | 99.999% | 99.99% |
| Billing unit | Processing Unit | RCU/WCU or on-demand | RU/sec | vCPU-hours |
| Serverless | — | Excellent (On-demand) | Excellent (Serverless) | Excellent |
| Track record | 10+ years (Google internal) | 10+ years | 10+ years | 2024- |
| Level | Spanner | DynamoDB | Cosmos DB |
|---|---|---|---|
| Strong / Linearizable | Excellent (default) | Yes (per request) | Yes |
| Bounded Staleness | — | — | Yes |
| Session | — | — | Yes (default) |
| Consistent Prefix | — | — | Yes |
| Eventual | — | Excellent (default, cheapest) | Yes |
| Service | Approx. monthly cost |
|---|---|
| Spanner (1000 PU = 1 node) | ~$200 (regional) |
| DynamoDB On-demand | ~$400 |
| DynamoDB Provisioned | ~$300 |
| Cosmos DB Provisioned (1000 RU/s) | ~$60 + data transfer |
| Cosmos DB Serverless | ~$280 |
| Feature | Spanner | DynamoDB | Cosmos DB |
|---|---|---|---|
| Secondary indexes | Excellent | GSI / LSI | Auto + Composite |
| Stream / CDC | Change Streams | Streams | Change Feed |
| PITR | 1 hour to 7 days | 35 days | 30 days |
| Time Travel | — | — | — |
| Vector Search | — | — | Excellent (since 2024) |
| Graph | Spanner Graph (2024) | — | Cosmos for Apache Gremlin |
| Use case | Recommended | Why |
|---|---|---|
| Financial transactions / banking | Spanner | Strong consistency + global + ACID |
| Serverless web / API | DynamoDB | On-demand + simple KV model |
| Mobile + web, multi-cloud | Cosmos DB | Multi-model + 60+ regions |
| E-commerce order management | Spanner | Transactions + global |
| IoT event logs | DynamoDB | Serverless + high write throughput |
| AWS PG-compatible + global | Aurora DSQL | Newly integrated into AWS |
| Knowledge graph | Spanner Graph / Cosmos Gremlin | Property Graph |
| Vector Search + document | Cosmos DB (2024) | Native vector support |
How do Spanner, DynamoDB, and Cosmos DB differ?
Spanner = NewSQL (SQL + ACID + global distribution), DynamoDB = serverless NoSQL KV/Document, Cosmos DB = multi-model NoSQL + SQL API. They are completely different beasts.
Is Aurora DSQL in the same category?
Yes (GA in 2024). It's AWS's Spanner-like NewSQL. The track record is still short, but it's already integrated into the Aurora ecosystem, bridging Spanner-style features into AWS.
Which offers strong consistency in a global setup?
Spanner = External Consistency (TrueTime-based, the strongest in the industry); Cosmos DB = 5 consistency levels (Strong, Bounded Staleness, etc.); DynamoDB = Eventual + Strong (within a single region).
What do the pricing models look like?
Spanner = Processing Units (100 PU = $0.90/h) + storage; DynamoDB = RCU/WCU or on-demand; Cosmos = RU/sec (Provisioned or Serverless).
How serverless are they?
DynamoDB On-demand bills only per request; Cosmos DB Serverless bills only per RU consumed; Spanner starts at a minimum 100 PU (~$70/month). DynamoDB is the most serverless of the three.
Can they handle multi-region writes?
Spanner: in multi-region setups every region can accept writes. Cosmos: native multi-region writes. DynamoDB: handled via Global Tables (last-writer-wins).
Which one for which workload?
Financial transactions → Spanner; serverless web/API → DynamoDB; multi-cloud + multi-model → Cosmos DB; AWS Aurora compatibility → DSQL.
What does a cost comparison look like?
For 1,000 reads/sec + 1,000 writes/sec + 100 GB of data: Spanner ~$200/month, DynamoDB On-demand ~$400/month, Cosmos Provisioned ~$300/month. The ranking can flip depending on the workload.
Related Articles / DB Comparisons
GCP vs AWS ストレージ・DB 徹底比較｜GCS/S3・BigQuery/Redshift・Spanner/DynamoDB (2026)
GCP と AWS のストレージ・データベースを徹底比較。Cloud Storage vs S3、BigQuery vs Redshift、Spanner vs DynamoDB / Aurora DSQL、Cloud SQL vs RDS、AlloyDB vs Aurora、Firestore vs DynamoDB、Bigtable vs DynamoDB を 2026 年最新版で網羅。
GCP vs Azure 完全比較｜Compute・Storage・AI・料金・認定 (2026)
Google Cloud と Microsoft Azure を徹底比較。Compute Engine vs Azure VM、BigQuery vs Synapse / Fabric、Cloud Run vs Container Apps、Gemini vs Azure OpenAI、ID 統合、ハイブリッド、認定試験、料金を 2026 年最新版で網羅。
Gemini vs GPT-4 vs Claude vs Llama 徹底比較｜LLM API 選び方・料金 (2026)
Google Gemini 2.0 / OpenAI GPT-4o / Anthropic Claude Opus 4 / Meta Llama 3.3 の徹底比較。性能 / コード / 推論 / マルチモーダル / 料金 / コンテキスト長 / GCP・AWS・Azure 経由利用を 2026 年最新版で網羅。
Vertex AI vs SageMaker vs Azure ML 徹底比較｜MLOps プラットフォーム選び方 (2026)
Google Vertex AI / AWS SageMaker / Azure ML の徹底比較。Gen AI 統合 (Gemini / Bedrock / Azure OpenAI)、AutoML、Pipelines、Feature Store、GPU/TPU、料金、認定試験を 2026 年最新版で網羅。
* Each product is the property of its respective trademark holder. Always check the official sites for the latest pricing.
Practice with certification-focused question sets
Visit the GCP Exam Prep PageNicheeLab Editorial Team
NicheeLab editorial team focused on data engineering and cloud certification learning. Content is structured around practical study needs and official exam domains.
Google Cloud Certification Roadmap (2026)
Choose your GCP certification path — Foundational, Associate...
CDL Cloud Digital Leader: Complete Exam Guide (2026)
Pass the Cloud Digital Leader exam — cloud business value, G...
GAIL Generative AI Leader: Complete Exam Guide (2026)
Pass the Generative AI Leader exam — Gemini, Vertex AI, Work...
Vertex AI Fundamentals for GCP Certs (2026)
Vertex AI basics every cert candidate needs — Workbench, Pip...
Associate Cloud Engineer (ACE): Complete Guide (2026)
Pass the Associate Cloud Engineer exam — Console, gcloud, pr...