🤖 AI Summary
To address replication latency, race conditions, and high network overhead induced by pointer-based designs in cross-region, high-throughput NoSQL database storage, this paper proposes the “Chunked Object” pattern. It decomposes oversized objects into ordered data chunks and stores them directly within a multi-region replicated key-value store (e.g., DynamoDB Global Tables), collocating data and metadata within a single consistency domain. This eliminates asynchronous replication between the database and external object storage, thereby significantly reducing cross-region p99 consistency attainment time. Evaluated in production under heavy load (>200K TPH, 1 MB/transaction), the approach demonstrates robust effectiveness and generalizability. It is applicable to any key-value store supporting per-item size limits and cross-region strong synchronization, offering a practical, system-agnostic solution for scalable, low-latency global data management.
📝 Abstract
Many managed key-value and NoSQL databases - such as Amazon DynamoDB, Azure Cosmos DB, and Google Cloud Firestore - enforce strict maximum item sizes (e.g., 400 KB in DynamoDB). This constraint imposes significant architectural challenges for applications requiring low-latency, multi-region access to objects that exceed these limits. The standard industry recommendation is to offload payloads to object storage (e.g., Amazon S3) while retaining a pointer in the database. While cost-efficient, this "pointer pattern" introduces network overhead and exposes applications to non-deterministic replication lag between the database and the object store, creating race conditions in active-active architectures.
This paper presents a "chunked-object" pattern that persists large logical entities as sets of ordered chunks within the database itself. We precisely define the pattern and provide a reference implementation using Amazon DynamoDB Global Tables. The design generalizes to any key-value store with per-item size limits and multi-region replication. We evaluate the approach using telemetry from a production system processing over 200,000 transactions per hour. Results demonstrate that the chunked-object pattern eliminates cross-system replication lag hazards and reduces p99 cross-region time-to-consistency for 1 MB payloads by keeping data and metadata within a single consistency domain.