Netflix AI Slashes Cassandra Read Latency Using Dynamic Repartitioning

Netflix's engineering team announced a novel method for managing wide partitions within Apache Cassandra, significantly improving read latency for its TimeSeries Abstraction platform. This platform handles petabytes of temporal event data, requiring millisecond query responses. The core innovation, termed Dynamic Partitioning, splits wide Cassandra partitions based on TimeSeries ID asynchronously and transparently, without necessitating application code changes. The system detects wide partitions on the read path using byte counting and Kafka events, prioritizing immutable partitions for splitting. Read requests are then efficiently routed to smaller child partitions using Bloom filters, which operate in single-digit microseconds, and a cached wide_row metadata lookup. Data integrity is maintained through checksums, retention of original partitions, and shadow comparisons via Data Bridge Spark checks.

This dynamic repartitioning strategy has resulted in a dramatic performance enhancement, reducing average read latency from seconds to low double-digit milliseconds. Tail latency has also seen substantial improvement, falling to approximately 200 milliseconds, even for partitions exceeding 500MB. Previously, wide partitions in Cassandra could lead to read timeouts, increased CPU utilization, and garbage collection pauses, particularly under the high read throughput demands of Netflix's TimeSeries servers. While scaling the Cassandra cluster is an option, Netflix sought a more intelligent solution than simply increasing hardware resources.

The TimeSeries Abstraction, which utilizes Apache Cassandra 4.x for its storage, was chosen for its throughput, latency, cost-effectiveness, and operational maturity. Time-series data is organized into partitions that group events by identifier and time range. As events accumulate, these partitions can grow excessively wide, leading to performance degradation. Dynamic repartitioning addresses this by splitting oversized partitions into smaller, more manageable child partitions. The applications continue to interact with the same logical partition, unaware of the underlying storage layout evolution, ensuring a seamless user experience while optimizing backend performance.