TIMESERIES TIERED STORAGE JOURNEY: KAFKA/FLINK STREAMS TO NATIVE CASSANDRA COLD READS

Summary from the TLDR Data newsletter

Netflix's TimeSeries Abstraction stores and queries multi-petabyte temporal datasets, tiering older, immutable slices from hot Cassandra into cheaper S3 storage. The earlier design used Kafka and Flink to stream Parquet to S3 and daily compaction, serving cold reads with roughly 500 ms p99 latency, but added operational cost. The later Cassandra-native cold tier reads directly from S3 backups, serves 15+ PB of compressed cold data, and improves p90 latency by about 30% while preserving the same query API.