Building DynamoDB at Scale

Key Takeaways

• DynamoDB is built to provide predictable low latency at any scale for applications
• The architecture uses horizontal partitioning and replication to achieve this, with automatic scaling and high availability
• The presentation covers three key aspects of the DynamoDB architecture:
  1. MemDS - an in-memory partition metadata store for fast lookup
  2. Warm Throughput - a feature to easily scale your table's capacity for spikes in traffic
  3. Strongly Consistent Global Tables - a new feature that provides global read-after-write consistency

MemDS - Partition Metadata Store

• DynamoDB needs to efficiently locate data partitions for millions of requests per second
• Using a traditional database for this metadata would not provide the required performance
• MemDS is an in-memory, distributed database designed to handle this lookup efficiently
• It uses versioning and eventually consistent updates from the storage nodes to maintain an accurate view of partition locations
• This allows request routers to quickly determine where to forward requests without introducing unpredictable latency

Warm Throughput

• Customers want to know how much traffic their DynamoDB tables can handle, especially for predictable spikes
• The old approach of provisioning capacity was complex and could lead to issues during scaling
• Warm Throughput allows setting a target throughput for a table or index, which DynamoDB will scale to meet
• This is independent of provisioned or on-demand billing mode, providing flexibility
• Warm Throughput can be set in advance of traffic spikes, ensuring the table is ready to handle the load

Strongly Consistent Global Tables

• Previous global table feature provided asynchronous replication and last-writer-wins conflict resolution
• New strongly consistent global tables use a multi-region log to guarantee global read-after-write consistency
• Writes are first recorded in the log, then asynchronously replicated to the table replicas
• Consistent reads use "heartbeats" to ensure they only return data that is fully replicated
• This allows use cases requiring strong consistency, like generating global unique sequence numbers

Takeaways for Developers

• Use hedging (sending duplicate requests) for predictable latency
• Design systems for constant work load to maintain availability
• Leverage eventual consistency where possible, but use synchronous caching carefully
• Maintain long-lived connections to DynamoDB for better performance
• Use the new Warm Throughput feature to scale for traffic spikes
• Consider strongly consistent global tables when you require cross-region consistency guarantees