On-Demand Features

Unlike pre-computed features that are materialized and stored, on-demand features execute transformation logic synchronously during the prediction request. The raw input data—such as a user's current session ID, geolocation coordinates, or real-time device context—is passed to the feature store, which applies the defined transformation and returns the computed feature vector immediately. This ensures the model always acts on fresh, context-specific data that would be impossible or impractical to pre-calculate for every possible input combination.

On-demand features rely on pure, stateless functions that take raw request data as input and return deterministic feature values. These functions are typically defined in Python or SQL and registered in the feature registry alongside pre-computed features. Key properties include:

• Idempotency: The same input always yields the same output
• No side effects: Functions do not write to external state
• Lightweight execution: Designed to complete in single-digit milliseconds to avoid adding latency to the inference path

A single inference request typically combines both pre-computed and on-demand features into a unified feature vector. The online store retrieves pre-materialized features—such as historical purchase aggregates or long-term preference embeddings—while simultaneously executing on-demand transformations for real-time context. The feature store's serving API orchestrates this parallel retrieval and computation, merging results before delivering the complete vector to the model endpoint.

On-demand features ingest data from the live request payload rather than historical data warehouses. Common input sources include:

• HTTP request headers: Device type, browser, IP-derived location
• Session metadata: Current page URL, time on site, scroll depth
• Real-time sensor data: GPS coordinates, accelerometer readings
• User-supplied parameters: Search query text, filter selections This architecture enables personalization based on immediate situational context that has no persistent historical record.

On-demand features are a critical tool for addressing the cold start problem in personalization systems. When a new user arrives with no historical interaction data, pre-computed features return null or default values. On-demand features compensate by extracting predictive signals from the current session context—such as referral source, landing page category, or device type—allowing the model to make informed predictions from the very first interaction without waiting for behavioral history to accumulate.

While on-demand features provide maximum freshness, they consume a portion of the end-to-end inference latency budget. Best practices include:

• Capping transformation complexity: Avoid heavy computation or external API calls
• Leveraging feature caches: Cache results for identical inputs within a time window
• Monitoring P99 latency: Track on-demand computation time separately from retrieval time Architects must balance the value of real-time context against the strict latency SLAs required for user-facing personalization.

The low-latency process of retrieving feature vectors and delivering them to a model endpoint. On-demand features are computed during this serving step rather than being retrieved from a pre-materialized cache. The serving infrastructure must route requests to the appropriate computation logic, often via a Serving API (REST or gRPC), and combine the results with pre-computed features from the Online Store to assemble the final feature vector.

Feature values computed incrementally on real-time event data using stream processing engines like Apache Kafka or Flink. Unlike on-demand features, which use data passed at request time, streaming features are pre-computed as events arrive and stored in the online store. They capture immediate user intent—such as session click count or items viewed in the last 5 minutes—and are refreshed continuously.

A one-dimensional array of numerical values representing the aggregated features for a specific entity at a specific point in time. When on-demand features are used, the final feature vector is a composite:

• Pre-computed features retrieved from the online store
• On-demand features computed at request time using client-supplied context This combined array is the direct input to the machine learning model for inference.

The low-latency database component of a feature store designed to serve pre-computed feature vectors during real-time prediction. On-demand features bypass the online store entirely—they are never materialized or persisted there. Instead, the online store supplies the static, pre-computed portion of the feature vector, while on-demand computation handles dynamic context like real-time geolocation, device type, or session-specific parameters passed in the inference request.

A metric defining the maximum acceptable age of a feature value before it is considered stale. On-demand features achieve zero staleness by definition—they are computed at the moment of request using the most current data available. This contrasts with batch features, which may have freshness measured in hours or days, and streaming features, which may have sub-second to minute-level freshness depending on pipeline latency.

A centralized metadata catalog that tracks feature definitions, schemas, lineage, and versions. On-demand features are registered here with their transformation logic and input parameters documented. This enables discovery and reuse across teams—a data scientist can find an existing on-demand feature definition, understand its required inputs, and incorporate it into a new Feature View without re-engineering the computation.