Feature store

A feature platform with a real-time feature store built in

Zipline.AI is a feature platform with a built-in feature store. Write a feature once in Python as a GroupBy (source and key, plus an aggregation like a 30-day count), and we generate the batch and streaming pipelines to backfill point-in-time-correct training data and serve values online. It runs on Chronon, the Apache 2.0 feature platform our founders built at Airbnb and still maintain. Stripe, Netflix, Uber and OpenAI run business-critical ML on Chronon.

What is a feature store?

Say your checkout model uses orders per user in the last 30 days. Training needs that count at the timestamp of millions of past orders, while live checkout needs the current number for one shopper in milliseconds. A feature store holds those values for both reads.

Trouble starts when two teams write that math twice: in Python for training, and months later in Java for production. Soon the two versions don't agree on window cutoffs, and the live model sees different numbers than it trained on.

What are real-time features?

Nightly batch jobs miss what happened five minutes ago. Fraud models need real-time features like card velocity, counting how many times a card was tried in the last ten minutes to block an attack as it happens.

Running that live takes a stream for new events paired with a batch job for training history. On Zipline.AI, you write the GroupBy once in Python and we generate both pipelines.

Feature store or feature platform?

A typical feature store only holds numbers you computed elsewhere, so your engineers still build the batch and streaming jobs. Zipline.AI is a feature platform that generates those pipelines from your Python code, backfilling training data and serving live features from its store.

CapabilityTypical feature storeZipline.AI
Online storage and servingStores values computed elsewhereOne key read per lookup
Batch and streaming pipelinesYou build and run bothGenerated from one Python GroupBy
Historical training dataJoins precomputed tablesBackfills point-in-time-correct history
Training and serving agreementTwo codebases must matchBoth use one definition

How Zipline.AI works

  1. 01

    Define

    Write a GroupBy in Python with a source and key, plus an aggregation like a 30-day order count.

  2. 02

    Backtest

    Zipline.AI generates a batch job to backfill point-in-time-correct training data over history so future events never leak into past labels.

  3. 03

    Ship

    The first batch upload seeds new windows from history, so a new 30-day feature serves correct values online within hours. Edits recompute only the feature groups that changed.

  4. 04

    Serve

    Each online lookup is one key read via GigaTile, built with OpenAI and Etsy. At 100,000 requests per second in our published comparison, Zipline.AI costs 5 to 9 times less to serve than Chalk and Databricks.

Part of your ML platform

Zipline.AI runs in your own AWS, GCP or Azure account. It reads from Snowflake, BigQuery, Databricks, Kafka and your services, writing to your own object store and key-value store.

Because a feature is a Python file and a backfill job, agents can run experiments on it too. At Stripe, Chronon sits under Shepherd (the feature platform behind Radar), where agents running experiments raised fraud recall by over 8%. See agentic improvement loops.

How it compares

Tecton
Tecton is now part of Databricks.
Databricks and Snowflake
Their feature stores are strongest at batch features inside their own warehouse.
Chalk
Chalk merges overlapping time buckets in its query servers on every request, while Zipline.AI does one key read per lookup.
Feast
Feast stores and serves features, and you build the batch and streaming pipelines yourself as separate code.

Read the published comparison, with a serving cost calculator

FAQ

What is a feature store?

It's the database layer between raw data and your models. Training pulls historical rows from your object store in bulk, and live services read current values for one key from your key-value store.

What is the difference between a feature store and a feature platform?

It comes down to who writes the pipelines. A feature store waits for numbers from batch and streaming jobs you built yourself, whereas Zipline.AI generates both jobs from one Python GroupBy to backfill training data and serve live values.

What are real-time features?

They are inputs computed from events seconds ago, like card velocity. Zipline.AI pairs a stream with a batch backfill and seeds new windows from history on the first upload, so a 30-day feature is live within hours.

What does point-in-time correct mean?

When backfilling training data for a past checkout, the model may only see feature values from before that exact second. If your join includes events from later that day, the future leaks into training and offline metrics lie to you.

Is Zipline.AI open source?

The engine underneath is Chronon, an Apache 2.0 feature platform our founders created at Airbnb and maintain today. Zipline.AI is our commercial platform around Chronon and includes the GigaTile layout built with OpenAI and Etsy.

Does Zipline.AI run in my own cloud?

Everything runs in your own AWS, GCP or Azure account. We read from Snowflake, BigQuery, Databricks, Kafka and your services, writing straight to your object store and key-value store.

How does Zipline.AI compare with Tecton, Feast and Databricks?

Tecton is part of Databricks now, and Databricks and Snowflake focus on batch features in their own warehouses, while Feast leaves you to write the batch and streaming pipelines as separate code. Zipline.AI generates both pipelines from one Python definition and costs 5 to 9 times less than Chalk and Databricks to serve at 100,000 requests per second.

Which companies use Chronon?

We built Chronon at Airbnb, and it runs business-critical ML at Stripe, Netflix, Uber and OpenAI. At Stripe it sits under Shepherd, the feature platform behind Radar, where agents running experiments lifted fraud recall by over 8%.

Proof of concept

A proof of concept takes 2 to 4 weeks in your own cloud, and our founders work on it with you.