The Storage Layer Streaming Deserved.
- Columnar streaming storage.
- Delta Joins that cut compute by 80%.
- Native Apache Iceberg tiering.
Apache Kafka® was built for event transport, not analytics. Apache Fluss™ was built for both. Proven at 3 PB scale, processing 40 GB/s with sub-second latency.
How can I use Apache Fluss on Ververica?
Apache Fluss is available on Ververica as a Private Preview. Fluss, an Apache top-level project, adds columnar streaming storage so Flink jobs read only the columns they need, Delta Joins that move join state out of Flink jobs, and automatic tiering of older data to Apache Iceberg or Apache Paimon. Union Read queries hot and lake data in one SQL statement.
Key Statistics
- 3PB
- Proven ScaleProduction deployment processing 40 GB/s
- 80%
- Less ComputeCPU & memory reduction via Delta Joins
- 90→1s
- CheckpointsFrom 90 seconds down to 1 second
- <1s
- Data FreshnessSub-second ingestion to query
The Problem
You're using Kafka for something it wasn't built for. Your Flink jobs read 100% of every Kafka message, even when they only need a few half the columns. Stream-to-stream joins pile up 100 TB+ of fragile state in RocksDB with 90-second checkpoints.
4 Key Capabilities
Column Pruning
Apache Arrow columnar format, read only needed columns. Big network savings.
Delta Joins
Read only changed rows and needed columns. 80% less CPU and memory, 1-second checkpoints.
Native Iceberg Tiering
Hot data on NVMe/SSD, cold data auto-tiers to Apache Iceberg on object storage
Union Read
One SQL statement across hot and cold data transparently
Production Proof
Technical Specification
| Category | Specification |
|---|---|
| Storage Format | Apache Arrow columnar + row-based log (dual format) |
| Query Interface | Flink SQL (streaming & batch), StarRocks, Spark |
| Data Freshness | Sub-second (single-digit millisecond typical) |
| Throughput | 40 GB/s proven in production (3 PB deployment) |
| State Management | Delta Joins: externalized from compute engine |
| Hot Storage | NVMe/SSD with primary key indexing (RocksDB) |
| Tiered Storage | Automatic tiering to Iceberg/Paimon on object storage |
| Column Pruning | Read only needed columns (Arrow columnar format) |
| Checkpoint Impact | 1-second checkpoints (down from 90s in comparable systems) |
| Compatibility | 100% Flink Table API / SQL compatible |
The Storage Layer for Real-Time
See how Fluss replaces Kafka for analytics workloads. Proven at 3 PB scale.