AWS re:Invent 2025 - Multi-Region strong consistency with Amazon DynamoDB global tables (DAT440)

Summary of AWS re:Invent 2025 - Multi-Region Strong Consistency with Amazon DynamoDB Global Tables

Introduction to Multi-Region Architectures

• Customers are building distributed applications that require data to be close to end-users across multiple regions
• This is driven by needs for low latency, compliance, and high availability
• Traditional single-region databases struggle with the challenges of multi-region replication

DynamoDB Global Tables

• DynamoDB Global Tables was created to support multi-region architectures
• Provides two replication models: Asynchronous and Strongly Consistent (Mercy)
• Each model has different trade-offs around performance, consistency, and failure handling

Asynchronous Global Tables

• Provides low-latency, predictable performance for writes
• Uses asynchronous replication to propagate writes to other regions
• Ensures monotonic reads on a per-item basis, but not across items
• Resolves write conflicts using last-writer-wins based on timestamps
• Handles regional failures and network partitions well, with data remaining available

Asynchronous Replication Process

• Writes are first durably recorded in the local region
• A replicator then propagates the write to other regions asynchronously
• Each region has its own replicator to maintain independence between failure domains
• Replication is monotonic per-item, but not guaranteed across items

Failure Modes

• If a region is completely isolated, it can still serve reads and writes locally
• During a network partition, regions can continue operating independently, with data converging when connectivity is restored
• Streams notifications only reflect writes that were actually applied, not replicated writes

Strongly Consistent (Mercy) Global Tables

• Provides strongly consistent reads from any region
• Writes are serialized across regions using a multi-region journal
• Ensures all writes are visible in the same order across all regions

Write Process

• Write requests are forwarded to a replication engine that generates a log entry and appends it to the multi-region journal
• The journal then notifies all regions to apply the log entry
• The region where the write originated returns success once the log entry is applied

Read Process

• Strongly consistent reads wait for a "read fence" log entry to be applied, ensuring all prior writes are visible
• This ensures serializable reads across regions

Failure Modes

• If a region is isolated, it cannot serve writes or strongly consistent reads until connectivity is restored
• During a network partition between regions, the isolated regions can still serve reads and writes, but with higher latency due to the longer replication path

Key Takeaways

• Asynchronous Global Tables provide high-performance, eventually consistent replication
• Strongly Consistent (Mercy) Global Tables offer serializable reads and writes across regions, with higher latency
• Customers should carefully consider their application's consistency, availability, and latency requirements when choosing between the two models
• Extensive testing, including chaos and scale testing, is crucial to ensure resilience in multi-region failure scenarios

Technical Details

• DynamoDB stores 3 copies of data in 3 availability zones within a region
• Replication is unidirectional between regions, with a dedicated replicator per region pair
• Conflict resolution uses last-writer-wins based on timestamps
• Strongly Consistent Global Tables use a multi-region journal to serialize writes
• Strongly Consistent reads wait for a "read fence" log entry to ensure all prior writes are visible

Business Impact

• Multi-region architectures enabled by DynamoDB Global Tables allow customers to build highly available, low-latency applications that serve users globally
• The choice between Asynchronous and Strongly Consistent replication models allows customers to optimize for their specific performance, consistency, and availability requirements
• Robust failure handling and testing ensures applications can withstand regional outages and network partitions, providing high business continuity

Examples

• Customers can use Asynchronous Global Tables to build applications with fast, predictable performance, even during regional failures
• Strongly Consistent Global Tables enable applications that require serializable reads and writes across regions, such as financial systems or healthcare applications
• The ability to choose between replication models allows customers to find the right balance between consistency, availability, and latency for their specific use cases