Real-time features, AI search, Agentic similarities
Varant Zanoyan and Nikhil Simha Raprolu of Zipline AI explain why traditional feature stores are the wrong abstraction. They detail the journey of Chronon, the open-source engine born at Airbnb and battle-tested at Stripe, which focuses on compute, orchestration, and real-time correctness to solve the hardest data engineering challenges in ML, from fraud detection to powering modern AI agents with features and embeddings.
The Core Problem: Feature Platforms Must Handle Compute
The central thesis behind Chronon is that the most difficult challenge in the MLOps lifecycle isn’t storage, but compute. Many platforms calling themselves “feature stores” are essentially key-value stores that leave the complex and critical task of feature computation to the user. This approach sidesteps the primary pain point.
Chronon was built to address this directly. It defines a feature platform as a system that manages the end-to-end process:
- Ingesting raw data.
- Computing features for both streaming and batch contexts.
- Serving features with low latency for online inference.
- Providing point-in-time correct features for offline model training.
- Ensuring consistency between the online and offline paths.
The difficulty lies in tasks like streaming aggregations and generating historically accurate training data. Attempting to create point-in-time correct features with standard SQL queries often results in an N^3 complexity problem, which “blows up very quickly.” This computational barrier is why so many teams abandon real-time ML and settle for less effective batch features.
Origins: Forged in the Fires of Fraud and Search at Airbnb
Chronon (initially named Zipline) was created at Airbnb out of necessity. While viewed as a travel company, Airbnb operates a massive internal payments company, processing large sums of money across international borders. This made it a prime target for fraud.
The nature of fraud is adversarial; attackers constantly evolve their methods. The ML team needed the ability to engineer, test, and deploy new features in hours, not weeks. The existing process was a “mind-bending” task, requiring engineers to manually stitch together Spark for batch processing, Flink for streaming, Airflow for orchestration, and online databases for serving—just to produce a single feature, or “signal.” Models for fraud and search often require thousands of such signals.
Chronon was built to abstract this complexity away. A user can define a feature in a SQL-like language, and the system automatically handles the underlying orchestration of batch, streaming, and serving infrastructure.
After its success in fraud, the platform was adopted by the Search and Personalization team—another high-volume, real-time use case. This success led to organic, bottom-up adoption by other product teams for non-ML use cases, such as calculating the real-time average star rating for a listing. This demonstrated the platform’s value as a general-purpose, real-time data engineering tool.
Open Source and Commercialization with Zipline AI
Before being released to the public, Chronon was battle-tested through a close, six-month collaboration with Stripe. This partnership validated its architecture at another large-scale company with demanding fraud-detection needs.
While the open-source engine is powerful, setting it up is challenging because it requires deep expertise in the complex big data technologies it abstracts (Spark, Flink, etc.). This led to the creation of Zipline AI, a company focused on productizing Chronon and making it easy to adopt. The goal is to offer an “open core” model that provides the benefits of a managed service without vendor lock-in.
Zipline AI is now expanding its focus “upstream” to solve foundational data infrastructure problems, such as simplifying the adoption of Apache Iceberg and establishing best practices for scalable data ingestion.
Powering the Next Wave: AI Agents and Embeddings
The principles of real-time feature engineering are directly applicable to modern AI systems. At Airbnb, Chronon was already being used to power customer support agents by providing them with real-time context about a user’s situation. This is precisely the challenge now known as “context engineering” for LLMs and agents.
The platform has evolved to provide native support for embeddings, blurring the line between traditional features and learned representations. It orchestrates complex graphs that combine structured features and embeddings to power sophisticated workflows:
- User Sequence Modeling: Aggregating a user’s last
Nactivities (e.g., clicks, views) into a window, feeding it into a model to generate a real-time user embedding for personalization. - Retrieval-Augmented Generation (RAG): For customer support, embedding a user’s query and retrieving the most relevant paragraphs from a large corpus of policy documents.
- Real-time Intent: Capturing the evolving intent of a conversation with a support agent as an embedding to guide the resolution process.
By solving the underlying data engineering challenges, Chronon and Zipline AI aim to make these advanced, real-time AI capabilities accessible to all companies, not just the tech giants who can afford to build the complex infrastructure from scratch.