Feature Store Architecture and How to Build One

15 min readNov 8, 2023

Learn about the Feature Store Architecture and dive deep into advanced concepts and best practices for building a feature store.

As machine learning becomes increasingly integral to business operations, the role of ML Platform Teams is gaining prominence. These teams are tasked with developing or selecting the essential tools that enable machine learning to move beyond experimentation into real-world applications. One such indispensable tool is a feature store. If you find yourself grappling with the complexities of data pipelines for your ML models, a feature store could be the solution you’re looking for. In this article, we aim to provide a comprehensive guide to understanding and implementing a feature store, layer by layer. Our goal is to help you make an informed decision on whether a feature store aligns with your needs.‍

Feature Store Architecture? ‘Why’ Before ‘How’

Let’s get real for a moment. You’re not building a feature store for the fun of it; you’re building it because you have real challenges that need real solutions. So, what’s driving you to consider a feature store in the first place? Here are some of the most compelling reasons we’ve heard:

Real-Time Feature Serving: Your machine learning models require features with low latency and the ability to scale. This isn’t just a nice-to-have; it’s essential for operational efficiency.

Standardization: You’re tired of the Wild West approach to feature pipelines. You want a standardized way to build, store, and manage features for all your ML projects.

Unified Data Pipelines: The days of maintaining separate pipelines for training and serving are over. You’re looking for a unified approach to reduce training/serving skew and make your life easier.

Feature Reusability and Efficiency: A centralized feature store not only makes it easier to share features across projects but also enhances discoverability, accuracy, and cost-efficiency. By having a single source of truth for your features, you avoid redundant calculations and inconsistent usage in models.

If any of these resonate with you, you’re in the right place. A feature store addresses these challenges head-on, providing a structured, scalable way to manage your features, from ingestion to serving. And the best part? It’s not a one-size-fits-all solution; it’s a framework that can be tailored to meet your specific needs and constraints.‍

So, what’s driving you to consider building a feature store in the first place?

Feature Store vs. Data Store vs. ETL Pipelines: Understanding the Nuances

As you navigate the landscape of data management for machine learning, you’ll encounter several key components, each with its own set of capabilities and limitations. While feature stores are the stars of this guide, it’s crucial to understand how they differ from traditional data stores and ETL (Extract, Transform, Load) pipelines. This will not only help you make informed decisions but also enable you to integrate these components seamlessly.‍

The Role of a Feature Store

A feature store is more than just a specialized repository for machine learning features; it’s an integral part of the ML ecosystem that manages the entire lifecycle of feature engineering. While we will delve into its architecture in the following sections, it’s important to understand that a feature store is not merely a data storage solution. It provides a comprehensive framework for feature creation, versioning, and serving in both real-time and batch modes. To gain a deeper understanding of what features are and why they are crucial in machine learning, you can read our article on What is a Feature Store.‍

Traditional Data Stores

In contrast, traditional data stores like databases or data lakes are more general-purpose. They are excellent for storing raw or processed data but lack the specialized capabilities for feature engineering and serving that feature stores offer. For instance, they don’t inherently support versioning of features or real-time serving with low latency. While you could build these capabilities on top of a traditional data store, it would require significant engineering effort, something that’s already taken care of in a feature store.

ETL Pipelines

ETL pipelines, on the other hand, are the workhorses of data transformation. They are responsible for extracting data from various sources, transforming it into a usable format, and loading it into a data store. While ETL pipelines are essential for data preparation, they are not designed to manage the complexities of feature engineering for machine learning. They are more like a one-way street, taking data from point A to point B, without the nuanced management and serving capabilities that feature stores offer.

The Interplay

Understanding the distinctions doesn’t mean choosing one over the others; it’s about leveraging each for what it does best. You could use ETL pipelines to prepare your raw data and load it into a traditional data store for initial storage. From there, the feature store can take over, ingesting this data, transforming it into valuable features, and serving them to your machine learning models. In this way, each component-ETL pipelines, traditional data stores, and feature stores-can play a harmonious role in your data ecosystem.‍

In the section that follows, we’ll take a comprehensive look at the architectural components that make a feature store more than just a data repository. We’ll explore how it serves as a robust framework for feature engineering, management, and real-time serving, all while ensuring scalability and reliability.‍

Feature Store Architecture: A Practical Guide for Building Your Own

Before we roll up our sleeves and get our hands dirty with the nuts and bolts of a feature store, let’s take a step back. Imagine you’re looking at a blueprint; it’s easier to build a house when you know where each room goes, right? The same logic applies here. A feature store design is essentially divided into three core layers, each with its own set of components and responsibilities:

Data Infrastructure Layer: Think of this as your foundation. It’s where raw data is ingested, processed, and stored. This layer is the bedrock upon which everything else is built. Key components include Batch and Stream Processing Engines, as well as Offline and Online Stores.

Serving Layer: This is your front door, the gateway through which processed features are made accessible to applications and services. It’s optimized for speed and designed for scale. Here, you’ll find RESTful APIs or gRPC services that serve your features.

Application Layer: Lastly, consider this your control room. It’s the orchestrator that ensures all other layers and components are working in harmony. From job orchestration to feature tracking and system health monitoring, this layer keeps the ship sailing smoothly.

Understanding this Feature Store architecture is crucial because it informs every decision you’ll make, from the tools you choose to the workflows you establish. So, keep this blueprint in mind as we delve deeper into each layer and its components. Trust us, it’ll make the journey ahead a lot less daunting.‍

The Data Infrastructure Layer: Where It All Begins

The Data Infrastructure Layer is the backbone of your feature store. It’s responsible for the initial stages of your data pipeline, including data ingestion, processing, and storage. This layer sets the stage for the more specialized operations that follow, making it crucial for the scalability and reliability of your entire system.‍

Batch Processing Engine

The batch processing engine serves as the computational hub where raw data is transformed into features. It’s designed to handle large datasets that don’t require real-time processing and prepares them for storage in the offline feature store.


  • Data Consistency: Consistency is key for maintaining the integrity of machine learning models. Ensure that features generated are consistent across different runs.
  • Versioning: Keep track of different versions of features. If a feature is updated or deprecated, this should be captured.
  • Concurrency: Plan for multiple batch jobs running simultaneously to ensure they don’t conflict with each other in the feature store.

Relevance to SDK Concepts

  • Data Sources: The engine is where raw data from various SDK-defined sources like SQL databases, flat files, or external APIs is ingested. Consider the latency and throughput requirements of these data sources when designing your SDK.
  • Feature Transformations: The engine executes SDK-defined transformations. Depending on the type of machine learning model, different transformations may be more suitable. For example, for classification models, you might consider label encoding, while for regression models, polynomial features could be useful.

Best Practices

  • Batch Sizing: Choose batch sizes that optimize both computational speed and system load. For time-series data, you might opt for daily or hourly batches.
  • Feature Validation: Implement checks to ensure that computed features meet quality and consistency standards, akin to a quality check before a dish leaves the kitchen.
  • Dependency Management: Manage the order in which features are computed, especially if one feature depends on another, similar to the steps in a recipe.

Highlighted Option

  • Apache Spark: Spark offers a distributed computing environment that’s highly scalable and fault-tolerant. It supports a wide range of data sources and formats, making it versatile for various feature engineering tasks. Its native support for machine learning libraries also makes it a robust choice for feature computation and storage in a feature store.‍

Stream Processing Engine‍

The Stream Processing Engine is like the “fast-food counter” of your data infrastructure, designed to handle real-time data processing needs. It processes data as it arrives, making it ideal for applications that require real-time analytics and monitoring.


  • Latency: Unlike batch processing, latency is a critical factor here. The system should be capable of processing data with minimal delay.
  • Scalability: As data streams can be highly variable, the system should be able to scale up or down quickly.
  • Data Integrity and Fixes: Mistakes happen, and sometimes incorrect data gets streamed. Your engine should not only handle out-of-order or late-arriving data but also be capable of correcting these errors either in real-time or through subsequent batch recalculations.

Relevance to SDK Concepts

  • Data Sources: This engine typically deals with real-time data sources defined in the SDK, such as Kafka streams, IoT sensors, or real-time APIs.
  • Feature Transformations: Stream-specific transformations like windowed aggregates or real-time anomaly detection can be executed here. For instance, if you’re working on a fraud detection system, real-time transformations could flag unusual transaction patterns.

Best Practices

  • State Management: Keep track of the state of data streams, especially if your features require data from multiple streams or have temporal dependencies.
  • Fault Tolerance: Implement mechanisms to recover from failures, ensuring that no data is lost and processing can resume smoothly.
  • Adaptive Scaling: Rather than imposing rate limits, focus on building a system that scales according to the demands of the incoming data streams.

Highlighted Option

  • Apache Spark Structured Streaming: We recommend Apache Spark Structured Streaming for its fault-tolerance, ease of use, and native integration with the Spark SQL engine. It allows for complex event-time-based window operations and supports various sources and sinks, making it versatile for real-time analytics and feature computation in a feature store. Its compatibility with the broader Spark ecosystem and DataFrame API makes it a robust choice for both batch and real-time data processing. The mature ecosystem, extensive community, and commercial support further solidify its standing as a go-to option for structured streaming tasks.‍‍

Offline Store‍

The Offline Store acts as your “data warehouse,” a secure and organized place where feature data is stored after being processed by the batch or stream engines. It’s designed to handle large volumes of data and is optimized for batch analytics.‍


  • Data Retention: Decide how long the data should be stored, considering both storage costs and data utility.
  • Accessibility: Ensure that the data is easily accessible for batch analytics but also secure.
  • Data Schema: Maintain a consistent schema to ensure that the data is easily interpretable and usable.

Relevance to SDK Concepts

  • Feature Sets: Feature sets are groups of features that share a common concept or relation. They are defined in the SDK and stored here. These could range from simple numerical features to more complex types like pre-processed text or images.
  • For example, if you’re building a recommendation engine, your feature set might include user behavior metrics like click-through rates, time spent on page, and purchase history. These features are related because they all contribute to understanding user preferences.
  • Feature Retrieval: This is your go-to for batch retrieval of features, often used for training machine learning models. Time-travel support is a part of this, allowing you to query features as they appeared at a specific point in time, which is useful for debugging or auditing.

Best Practices

  • ACID Transactions: Implement ACID (Atomicity, Consistency, Isolation, Durability) transactions to ensure data integrity.
  • Indexing: Use indexing to speed up data retrieval, especially for large datasets.
  • Data Validation: Before storing, validate the data to ensure it meets the quality and consistency requirements.

Highlighted Option

  • S3 with Delta or Iceberg files: Think of this as your high-security, climate-controlled warehouse. These file formats offer ACID transactions, scalable metadata handling, and unify streaming and batch data processing, making them a robust choice for an offline store.‍‍

Online Store‍

The Online Store is akin to your “retail shop,” designed for low-latency access to feature data. It’s optimized for quick reads and is the go-to place for real-time applications.


  • Latency: Low-latency is crucial here; data should be retrievable in milliseconds.
  • High Availability: The store should be highly available to meet the demands of real-time applications.
  • Scalability: As the number of features or the request rate grows, the system should scale seamlessly.

Relevance to SDK Concepts

  • Feature Retrieval: This is the primary source for real-time feature retrieval, often for serving machine learning models in production.
  • On-Demand Feature Computation: If the SDK supports it, some lightweight feature computation can also be done here in real-time.

Best Practices

  • Data Partitioning: Use partitioning strategies to distribute data across multiple servers for better performance.
  • Caching: Implement caching mechanisms to speed up frequent data retrievals.
  • Consistency: It’s crucial to maintain data consistency between the online and offline stores. This is especially important if both stores are updated simultaneously. Transactional integrity and recoverability are key here. For instance, if data is successfully written to the offline store but fails to write to the online store, you’ll need a robust mechanism to handle such discrepancies and recover gracefully.

Highlighted Option

  • Redis Caching: Redis is an open-source, in-memory data structure store that provides ultra-fast read and write operations. Its low-latency and high-throughput capabilities make it an excellent choice for serving features in real-time machine learning applications. With various data structures and atomic operations, Redis offers the flexibility to design efficient feature stores tailored to specific needs.‍

The Serving Layer: API-Driven Feature Access

The Serving Layer is your “customer service desk,” the interface where external applications and services request and receive feature data. It’s optimized for high availability and low latency, ensuring that features can be served quickly and reliably.‍


  • API Design: The APIs should be designed for ease of use, with clear documentation and versioning.
  • Load Balancing: Distribute incoming requests across multiple servers to ensure high availability and low latency.
  • Security: Implement authentication and authorization mechanisms to control access to feature data.

Relevance to SDK Concepts

  • Feature Retrieval: This layer is responsible for serving features to external applications, usually through RESTful APIs or gRPC services defined in the SDK.
  • On-the-Fly Computations: In addition to serving precomputed features, this layer can also perform lightweight computations in real-time as per the SDK’s capabilities. For example, if you’re serving features for a recommendation engine, you might need to calculate the “popularity score” of an item based on real-time user interactions. This score could be computed on-the-fly at the Serving Layer before being sent to the application.

Best Practices

  • Rate Limiting: Implement rate limiting to prevent abuse and ensure fair usage.
  • Monitoring: Keep track of API usage, errors, and latency for ongoing optimization.
  • Caching: Use caching mechanisms to speed up frequent data retrievals, much like a well-organized customer service desk that quickly retrieves common forms or information for customers.

Highlighted Option

  • Kubernetes: For a robust Serving Layer, we recommend a Kubernetes Cluster with the managed service provider of your choice. Complement this with Prometheus for real-time monitoring of system metrics and Kafka for effective rate limiting. When you have a high volume of incoming requests, these services can queue them up and feed them to the serving layer at a controlled rate. This helps in preventing the system from being overwhelmed and ensures that resources are used optimally. Rate limiting is especially important to prevent abuse and ensure fair usage of resources‍

The Application Layer: The Control Tower

The Application Layer serves as the orchestrator for your feature store. It manages the data pipelines, keeps track of features and their metadata, and monitors the system’s health. This layer ensures that all components of your feature store work in harmony, making it key for the system’s overall performance and reliability.‍

Job Orchestrator

‍The Job Orchestrator is the “conductor of the orchestra,” coordinating various components to work in harmony. It orchestrates your data pipelines, ensuring that tasks are executed in the correct sequence and managing dependencies between them.‍


  • Workflow Design: Define clear Directed Acyclic Graphs (DAGs) or workflows that outline the sequence and dependencies of tasks.

Relevance to SDK Concepts

  • Feature Sets: The orchestrator triggers the computation and storage of feature sets defined in the SDK.

Best Practices

  • Idempotency: Design tasks to be idempotent, meaning they can be safely retried without side effects, akin to a conductor who can restart a musical piece without causing confusion.
  • Integrated Monitoring and Logging: Incorporate monitoring dashboards and job logs into the feature store UI for rapid debugging without compromising access. This allows for a centralized view of job performance and issues, facilitating quicker resolution. Monitoring could include tracking the ‘freshness’ of data, latency, and error rates. For example, if you’re ingesting real-time stock prices, you might set an alert if data hasn’t been updated in the last 5 minutes.
  • Data Validation and Alerting: While it’s challenging to ensure the absolute correctness of computed data, implementing data validation checks and alerting mechanisms can help. For instance, if an ETL job is supposed to aggregate sales data, a sudden 50% drop in sales might trigger an alert for manual review.

Highlighted Option

  • Airflow: Airflow is an open-source platform designed to programmatically author, schedule, and monitor workflows. Its rich set of features and extensibility make it a robust choice for orchestrating complex data pipelines, including those in MLOps. With native support for defining task dependencies and scheduling, Airflow provides a comprehensive solution for workflow management. It also offers integration points for monitoring and alerting through tools like Prometheus and Grafana, or alerting services like PagerDuty.‍

Feature Registry (Metadata Store)

‍The Feature Registry serves as the “library catalog” of your feature store, maintaining metadata about each feature, its lineage, and other attributes. It’s the backbone that supports CRUD operations for metadata and offers feature lineage tracking.‍


  • Metadata Schema: Define a clear schema for metadata, including feature names, types, and lineage information.
  • Searchability: Ensure that features can be easily searched and retrieved based on their metadata.
  • Versioning: Implement versioning for features to track changes over time.

Relevance to SDK Concepts

  • Feature Sets: Metadata about feature sets defined in the SDK is stored here. This includes details like feature types, default values, and data sources.

Best Practices

  • Data Lineage: Maintain a record of the lineage of each feature, showing its journey from source to serving. This is akin to a library catalog that not only lists books but also shows their origins and how they arrived at the library.
  • Access Control: Implement fine-grained access control to restrict who can view or modify feature metadata.
  • Audit Trails: Keep logs of who accessed or modified features, similar to a library’s borrowing history.

Highlighted Option

  • PostgreSQL with Feast: This relational database offers robust capabilities for storing metadata. When used in conjunction with Feast, a feature store framework, you get additional benefits like feature lineage tracking and easy integration with data pipelines.‍

The Control Plane‍

The Control Plane is the “air traffic control tower” of your feature store, overseeing all operations and ensuring they run smoothly. It serves as the UI for data drift monitoring, access controls, and other management features.‍

Best Practices

  • Data Drift and Skew Monitoring: Implement algorithms to detect data drift and skew, which are crucial for maintaining the integrity of machine learning models.
  • Alerting: Set up alerting mechanisms for critical events or anomalies, with integrations such as Slack Webhooks, Opsgenie, etc.
  • Audit Logs: Maintain logs for all operations, providing a clear history of changes and access, much like an air traffic control log.

Highlighted Option

  • Serving Layer’s Kubernetes: Given that we recommend Kubernetes for the Serving Layer, it makes sense to use the same cluster for the Control Plane as well. This offers a cohesive, scalable, and cost-effective management experience, and simplifies the architecture by reducing the number of services you need to manage.‍

Wrapping Up

Building a feature store is not for the faint of heart; it’s a complex endeavor that requires a deep understanding of both your data and the tools at your disposal. But here’s the good news: you don’t have to go it alone. From managed services to open-source projects, there are resources out there to help you on how to build a feature store. The key is to start with a solid foundation, and that begins with understanding the feature store architecture. We’ve walked you through the Data Infrastructure, Serving, and Application Layers, demystifying the components that make them up. Remember, a feature store is more than just a sum of its parts; it’s an ecosystem. And like any ecosystem, balance is key.‍

So as you embark on this journey, keep your eyes on the horizon but your feet firmly on the ground. After all, the future of machine learning is not just about algorithms; it’s about features — how you store them, manage them, and serve them. And that’s a future worth building for.‍

Originally published at https://www.qwak.com.




A fully managed AI platform that unifies ML engineering and data operations.