Megabytes in milliseconds: How FireTV uses parallel queries and vertical partitioning to serve millions of customers in Amazon DynamoDB

This article describes how FireTV redesigned their DynamoDB data model for Continue Watching using vertical partitioning and hash-prefixed sort keys to overcome item size constraints and improve performance.

• Original single-item design approached 400 KB limit; doubling data would exceed it for many customers
• Large items caused throttling and high write costs (50 WCUs per update)
• Vertical partitioning breaks large items into smaller items, reducing write costs to 1 WCU per update
• Hash-prefixed sort keys enable parallel segmented queries to avoid 1 MB page limit pagination
• Parallel queries return in single round trip per segment, totaling 8 ms average read latency
• Segment count tunable via client-side configuration without schema migration
• Tradeoff: loses efficient fine-grained access patterns; requires GSI for per-provider queries
• Validation used parallel table writes, backfill, shadow reads, and EMR comparison of billions of records

The design eliminates size constraints, scales to tens of thousands of entries per profile, and maintains consistent low-latency reads regardless of profile size.