How Databricks Feature Store serves features with sub-second freshness

Databricks has introduced sub-second feature freshness in its Feature Store, addressing a critical gap in real-time machine learning where models require immediate access to the latest data. Traditional batch processing introduces minutes to hours of lag, forcing data scientists to build custom streaming pipelines for fresh signals. The new capability leverages Spark Real-Time Mode (RTM) to process streaming data concurrently, reducing latency from seconds to milliseconds. This eliminates the need for separate infrastructure, integrating batch and streaming pipelines under a single framework.

The system processes incoming events through a stateful pipeline that updates rolling aggregations—such as transaction sums over a 10-minute window—in real time. Each event is routed to a local RocksDB instance, which maintains running totals with automatic window expiration. Updated feature values are then written to Lakebase, an online storage layer optimized for low-latency reads. For fraud detection, this enables models to access the most recent transaction sums alongside historic baselines, improving decision-making within milliseconds of an event occurring.

Underlying the feature freshness is Spark RTM, a new execution mode that processes streaming data in parallel rather than in discrete microbatches. This approach reduces latency by eliminating batch boundaries and amortizing checkpointing costs over longer intervals. RTM maintains exactly-once processing guarantees while allowing pipelines to recover from failures by replaying at most five minutes of data. The architecture also decouples compute and storage, enabling autoscaling for high-throughput inference workloads without manual cluster management.

The final component is Databricks Model Serving, which retrieves features from Lakebase during inference with minimal latency. Features are automatically joined with model requests, eliminating the need for custom lookup logic. The platform also addresses training data generation for streaming features by maintaining an offline copy of Kafka data, enabling point-in-time accurate joins for model training. Governance is streamlined through Unity Catalog, where features are first-class objects with full lineage and access controls.