AWS China Region: 5 Critical Gaps Beyond the Service Catalog (Live Test Data)

Five Dimensions of AWS China Region Gaps: From Catalog Checkmarks to Production Reality

This analysis is based on live read-only API testing in AWS Beijing and Ningxia regions (data collected 2026-10-07/09). The author moves beyond the service catalog (120+ services) to measure five practical dimensions: version parity, functional completeness, quota limits, developer experience, and pricing.

Version Parity: Core Services Mostly Synced

Testing 18 common engines and components across three partitions (global, Beijing, Ningxia) revealed that version lag is not the primary issue. 14 of 18 services are fully version-aligned. Open-source databases track closest: MySQL, MariaDB, DocumentDB, and even Valkey (released 2024) are current. The claim that "China region database versions lag" is false per this test.

Three notable gaps remain:

Aurora PostgreSQL trails by two minor versions — teams needing latest features must verify version numbers before migration.

Amazon MQ is the only service with a major-version gap: RabbitMQ and ActiveMQ each lag one major version, impacting message-heavy architectures.

Neptune 1.4.8.1 is available only in Ningxia; Beijing remains on 1.4.8.0 — a real intra-China difference.

EKS Auto Mode (launched late 2024) is available in China. The API does not expose an explicit switch; the author detected it via create-cluster with computeConfig parameter, which returned validation errors referencing Auto Mode integrity rules — confirming the backend logic is in place.

Overall, China region offers ~1/3 of global service count, but nearly all core-layer services (EKS, RDS, Aurora, ElastiCache, OpenSearch) are version-current. The missing services fall into two categories: policy-sensitive (SES, SMS) and emerging AI (Bedrock, Q, CodeWhisperer). For policy-sensitive gaps, cheaper local alternatives exist; for AI inference, domestic providers like DeepSeek and MiniMax offer OpenAI/Claude-compatible APIs, making migration less challenging.

Functional Completeness: Grayed-Out Switches Behind the Checkmark

Functional gaps appear in two forms: entire product lines absent (Bedrock, GPU training instances) and key sub-features disabled in otherwise-available services.

AI, Edge & Compute

Critical signal: CloudFront in Ningxia will be deprecated on 2027-05-31 (official notice). This is not a feature gap but a service exit; teams must plan migration to domestic CDNs.

AI compute ceiling: strongest instance is g5.48xlarge (8× A10G, 24 GB VRAM each) . Models ≤24 GB VRAM can run inference; training/fine-tuning 70B+ models is currently impossible. Managed AI APIs are entirely absent: Bedrock, Rekognition, Textract, Translate, Polly, Q, CodeWhisperer are unavailable in Beijing. SageMaker is the only fully available AI service , but it hits the same hardware ceiling — no large-VRAM training instances.

Data & Governance

Conclusion: Data layer can be migrated almost blindly. DynamoDB fully parity; ElastiCache includes Valkey; DocumentDB lacks only previous-generation instances.

Two traps:

Observability gap: Metrics, synthetics, SLOs present, but RUM (Real User Monitoring) explicitly unavailable — frontend monitoring teams need alternatives.

Governance cannot be inherited: Global Organizations and Identity Center do not extend to China. However, China region can run a full, non-crippled Organizations + Identity Center stack with new-feature APIs working. Multi-account strategy, SCPs, SSO must be rebuilt from scratch. Architecture can be copied; governance cannot be inherited.

Quotas: The Landmines Found During Load Testing

Service Quotas API returns two layers: default (AWS documented defaults) and applied (account-specific values). EC2 applied quotas grow with usage and are partition-agnostic. ECS applied layer is empty because nearly all ECS quotas are "non-adjustable" — only adjustable quotas appear in applied layer.

After filtering account noise, the only true partition difference in quotas is that EKS has no Service Quotas integration in China — both API layers return "service unavailable", and EKS is absent from the Console quota list. Node-count and cluster-count limits can only be obtained via support ticket. IaC that references EKS quota values for guardrails will fail.

Key reminder: applied values are polluted by account history; only default values reflect partition policy. Using the wrong layer misreads account differences as partition differences.

Notable non-adjustable defaults (Console quota dashboard: https://console.amazonaws.cn/servicequotas/home/dashboard):

Over half of Console quotas are marked "non-adjustable" — hitting them forces architecture changes, not support tickets.

Developer Experience: Docs, Images, Toolchain

Dedicated docs site: https://docs.amazonaws.cn — not fully synced with global docs; verify you are reading the China version when behavior mismatches documentation.

Container image pulls: EC2/EKS nodes inside China cannot reach Docker Hub, GitHub Container Registry, or Kubernetes official registries. Workaround: pull from self-hosted mirrors/proxies outside China. GitHub.com itself is largely reachable; cloning small-to-medium repos is acceptable. Build images locally in China region.

AWS CLI: Works seamlessly. Configure two profiles (one for China, one for global) on the same machine. China endpoints are separate domains but transparent to CLI; direct connectivity from mainland to China region services is solid.

Domain & DNS: China region does not support domain registration. Buy from a domestic registrar, complete ICP filing, then use the domain in China region. Unfiled domains resolve via DNS but websites won't load.

No CloudShell: Unlike global, China region lacks CloudShell. Engineers accustomed to browser-based terminals must bring their own local toolchain.

Cost: Beijing vs. Ningxia Pricing Inversion

Counter-intuitive finding: Beijing–Ningxia pricing is inverted. Beijing is 23–48% more expensive than global; Ningxia prices are generally equal to or lower than global. Example: g5.12xlarge in Ningxia ≈ $5.32/hr, 6% cheaper than US-East. A 30% price spread between two regions in the same country has no parallel in global partitions.

Ningxia pricing appears designed to steer compute workloads there. For latency-insensitive AI inference, placing GPU workloads in Ningxia saves ~1/3 vs Beijing and beats US-East. Region selection within China deserves its own cost model.

Pre-Launch One-Hour Checklist (Verified Commands)

Check versions: Run aws eks describe-cluster-versions and aws rds describe-db-engine-versions; compare target region against current region.

Check instance types: Run aws ec2 describe-instance-type-offerings; confirm every instance type your architecture requires exists in China. For GPU workloads, verify specific chip models.

Check service availability: On docs.amazonaws.cn, verify each service your architecture depends on is supported in China.

Check quotas: First run aws service-quotas get-service-quota; if empty, run get-aws-default-service-quota. EKS in China returns "service unavailable" — that error is the answer.

Check pricing: Use Pricing API to compare all three partitions; Beijing–Ningxia delta may change deployment decisions.

From Gaps to Decisions

AWS China is not a "lite" version of global — it is a distinct target platform requiring independent evaluation.

Strengths: Version sync on 14/18 core services; EKS Auto Mode present; default quotas aligned (only EKS quota API missing); Ningxia pricing beats US-East; data layer near-frictionless; governance stack fully rebuildable.

Hard limits: AI compute frozen at 2021-level (A10G 24 GB); managed AI APIs entirely absent; CloudFront exit scheduled.

The migration decision hinge is not "does AWS China work?" but "which layer does your architecture press on?" Core compute, databases, containers, observability — China is a credible, cost-friendly target. AI workloads — only SageMaker self-managed, capped by available silicon.

This article will be updated with further live tests. For teams doing China-region architecture review, Chinaready's assessment service can operationalize this five-dimension checklist; for catalog-level parity list, see Chinaready Parity Checklist (https://martinliu.cn/chinaready/); for "overseas site slow in China" troubleshooting, see this guide (https://martinliu.cn/blog/why-overseas-sites-slow-in-china/).