DynamoDB DAX & Global Tables

DAX (DynamoDB Accelerator) is a write-through caching service that is API-compatible with DynamoDB, so applications use the DAX SDK client and point it at the cluster with minimal code change. It serves cached reads in microseconds instead of the single-digit milliseconds a direct table read takes, which suits read-heavy workloads with repeated key access such as product catalogs, gaming leaderboards, and configuration lookups. DAX maintains two caches: an item cache for GetItem and BatchGetItem results, and a query cache for Query and Scan results. Cached entries expire based on a configurable Time To Live (TTL).

DAX speeds up eventually consistent reads, because those can be served from the cache. A strongly consistent read must reflect the most recent write, so DAX passes it through directly to the table and returns table-latency results. Write-heavy workloads gain little, because every write goes through DAX to the table (write-through) and still incurs write latency. Workloads with low cache-hit ratios (mostly unique, one-time key access) also see little benefit, since few reads are served from cache.

A DAX cluster runs inside your VPC and consists of one or more nodes. One node is the primary that handles writes and propagates them to the cluster, while the others are read replicas that serve cached reads and provide availability. For production, run a cluster of at least three nodes spread across multiple Availability Zones, so the loss of one AZ or node does not take the cache offline. The application connects to a single cluster endpoint, and the DAX client distributes reads across nodes. Access is controlled through IAM and security groups, and DAX integrates with encryption at rest and in transit.

A Global Table is a single table replicated across multiple AWS Regions, where every replica is both readable and writable (active-active). DynamoDB automatically propagates each write to every other Region, so an application can read and write to the nearest Region for low local latency and continue operating if one Region becomes unavailable. This serves global user bases and supports a multi-Region disaster-recovery strategy with low recovery time and recovery point objectives. There is no extra charge for the Global Tables feature itself, but you pay for the storage, reads, writes (billed as replicated write request units), and cross-Region data transfer in each Region.

Because every Region accepts writes, two Regions can update the same item at nearly the same time. DynamoDB resolves this with a last-writer-wins policy: the write with the most recent timestamp is kept and the other is discarded. Design applications to avoid concurrent multi-Region writes to the same item when possible (for example, route a given user or partition key to a home Region) so that last-writer-wins does not silently drop data.

DynamoDB Streams is an ordered, time-ordered log of item-level changes (inserts, updates, deletes) in a table, retained for 24 hours. Global Tables use Streams as the replication mechanism: each write is captured in the stream and asynchronously delivered to the other Regions. Because of this, Streams must be enabled on a table before it can join a Global Table (the newer version uses NEW_AND_OLD_IMAGES). Streams also serve other use cases on their own, such as triggering AWS Lambda functions for event-driven processing, change-data-capture, and aggregation.

An eventually consistent read may return data that does not yet reflect a very recent write, because the read can be served from a replica that is slightly behind. A strongly consistent read always returns the latest committed write within a single Region, at the cost of higher latency and not being cacheable by DAX. Two layers of eventual consistency apply: within one Region, DynamoDB reads are eventually consistent by default; across Regions, Global Tables replicate asynchronously, so a write in one Region appears in the others usually within a second but not instantly. Strong consistency is a single-Region guarantee and does not extend across Regions.