Netflix AI Team Cuts Cassandra Wide-Partition Read Latency from Seconds to Milliseconds

Netflix's AI team faced multi-second read latency with Cassandra wide partitions in their TimeSeries Abstraction layer. They implemented two complementary methods: time-sliced repartitioning for table-level future partitioning, and dynamic partitioning that detects and splits oversized partitions on the read path. Dynamic partitioning uses byte counting and Kafka detection, performs checksum validation after splitting, and uses Bloom filters to route reads to parallel sub-partitions. Average read latency dropped from seconds to low double-digit milliseconds, even for partitions over 500MB. This approach allows engineers managing large-scale time-series data on Cassandra to automatically handle data skew without schema changes, reducing hotspot impact and enabling stable low latency for real-time analytics.