Netflix migrates 15PB of data to Cassandra-native S3 cold storage
Netflix has detailed its transition to a tiered storage architecture for its TimeSeries platform, migrating 15PB of data from hot Cassandra clusters to S3-backed cold storage. The project utilizes custom Cassandra Memtables and ReadOnlyCompactionStrategy (ROCS) to maintain low-latency query performance while significantly reducing infrastructure overhead and storage costs.
Key Takeaways
- Migrated 15PB of compressed data from high-cost hot Cassandra clusters to S3-backed cold storage.
- Developed ReadOnlyCompactionStrategy (ROCS) to align SSTables with token ranges, improving p90 latency by 30%.
- Leveraged plugable Memtables in Cassandra 4.1 to bypass streaming overhead and read directly from existing S3 backups.
- Reduced storage utilization on target clusters from 80% disk usage to just 7% during the rollout phase.
- Maintained p99 query latency under 500ms for historical data despite moving off local SSDs.
Why It Matters
This architectural shift demonstrates a viable path for streaming platforms to solve the conflict between long-term data retention and infrastructure costs. By decoupling the storage engine from physical disks and reading directly from backup object stores, Netflix has effectively eliminated the 'double-pay' scenario for data durability and availability. For the broader ecosystem, this validates S3 as a primary read-tier for high-scale databases rather than just a disaster recovery sink. Engineering leaders should watch for the potential open-sourcing of the ReadOnlyCompactionStrategy and backup Memtable implementations as standard Cassandra features.
Additional Context
The move to Cassandra-native cold storage is part of a multi-year efficiency push at Netflix to manage a database fleet that has tripled in size since 2022. Per reporting from AxonOps and Community Over Code 2025, Netflix now operates over 2,500 Cassandra clusters across 58,000 nodes, processing 38 million requests per second. As data volumes grew from 12PB to over 20PB in just three years, the company has increasingly moved toward 'Data Abstraction Layers' to shield developers from the underlying storage complexity while optimizing for costs that reportedly exceed $1 billion in annual AWS spend.
Technically, this migration follows the General Availability of Netflix’s 'Data Bridge' in March 2024, a unified control plane designed to standardize data movement across its ecosystem. While earlier iterations of TimeSeries storage relied on Flink and Kafka for streaming to S3, the new native approach addresses what Netflix engineers described in May 2026 as the 'monolithic design' and 'fragile metadata dependencies' of legacy systems like Casspactor. By reading directly from S3-based backup manifests, Netflix has reduced the compute overhead and data transfer costs associated with traditional ETL pipelines.
This shift also reflects broader cloud industry trends where egress fees and storage density have become primary blockers to scale. Per a 2025 Dimensional Research report, 95% of organizations experience unexpected cloud storage charges, leading many to adopt tiered storage policies. Netflix's implementation of ROCS and pluggable Memtables effectively bypasses these constraints by repurposing disaster recovery assets into active read-only storage tiers, a strategy that is now being explored by other high-volume data users like Bloomberg and Akamai.
Read full article at netflixtechblog.medium.com
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source