Chalk home page
  1. Overview
  2. What is Chalk?

The Data Platform for Machine Learning

Tired of Spark? So are we.

Chalk is a feature store that enables innovative machine learning teams to focus on building the unique products and models that make their business stand out. Behind the scenes, Chalk seamlessly handles data infrastructure with a best-in-class developer experience. Here’s how it works.


Chalk makes it simple to develop feature pipelines for machine learning. Define Python functions using the libraries and tools you’re familiar with instead of specialized DSLs. Chalk then orchestrates your functions into pipelines that execute in parallel on a Rust-based engine and coordinates the infrastructure required to compute features.


To get started, define your features with Pydantic-inspired Python classes. You can define schemas, specify relationships, and add metadata to help your team share and re-use work.

from datetime import datetime
from chalk.features import features, DataFrame

class Transaction:
    id: int
    amount: float
    user_id: ""
    user: User

class User:
    id: int
    full_name: str
    nickname: str | None
    email: str | None
    birthday: datetime
    credit_score: float
    datawarehouse_feature: float
    transactions: DataFrame[Transaction]


Next, tell Chalk how to compute your features. Chalk ingests data from your existing data stores, and lets you use Python to compute features with feature resolvers. Feature resolvers are declared with the decorators @online and @offline, and can depend on the outputs of other feature resolvers.

Resolvers make it easy to rapidly integrate a wide variety of data sources, join them together, and use them in your model.

pg = PostgreSQLSource()

def get_user(uid: -> Features[User.full_name,]:
    return (


Once you’ve defined your features and resolvers, Chalk orchestrates them into flexible pipelines that make training and executing models easy.

Chalk has built-in support for feature engineering workflows — there’s no need to manage Airflow or orchestrate complicated streaming flows. You can execute resolver pipelines with declarative caching, ingest batch data on a schedule, and easily make slow sources available online for low-latency serving.


Many data sources (like vendor APIs) are too slow for online use cases or charge a high dollar cost-per-call. Chalk lets you optimize latency and cost by defining declarative caching policies that are well-integrated throughout the system. You no longer have to manage data sources such as Redis, Memcached, or DynamoDB, or spend time tuning cache-warming pipelines.

Add a caching policy with one line of code in your feature definition:

class ExternalBankAccount:
  balance: int
  balance: int = feature(max_staleness="1d")

Optionally warm feature caches by executing resolvers on a schedule:

from chalk import online

def fn(id: -> User.credit_score:
  return redshift.query(...).all()

Or override staleness tolerances at query time when you need fresher data for your models:

from chalk.client import ChalkClient

    outputs: [User.fraud_score],
    max_staleness: { User.fraud_score: "1m" }

Batch ETL ingestion

Chalk also makes it simple to generate training sets from data warehouse sources — join data from services like S3, Redshift, BQ, Snowflake or other custom sources with historical features computed online. Specify a cron schedule on an @offline resolver, and Chalk will automatically ingest data with support for incremental reads:

def fn() -> Feature[, User.datawarehouse_feature, User.updated_at]:
  return redshift.query(...).incremental()

Chalk makes this data available for point-in-time-correct dataset generation for data science use cases. Every pipeline has built-in monitoring and alerting to ensure data quality and timeliness.

Reverse ETL

When your model needs to use features that are canonically stored in a high-latency data source (like a data warehouse), Chalk’s Reverse ETL support makes it simple to bring those features online and serve them quickly.

Add a single line of code to an offline resolver, and Chalk constructs a managed reverse ETL pipeline for that data source:


Now data from slow offline data sources is automatically available for low-latency online serving.

Deploying and querying

Once you’ve defined your pipelines, you can rapidly deploy them to production with Chalk’s CLI:

chalk apply

This creates a deployment of your project, which is served at a unique preview URL. You can promote this deployment to production, or perform QA workflows on your preview environment to make sure that your Chalk deployment performs as expected.

Once you promote your deployment to production, Chalk makes features available for low-latency online inference and offline training. Chalk uses the exact same source code to serve temporally-consistent training sets to data scientists and live feature values to models. This re-use dramatically shortens development time and ensures that feature values from online and offline contexts match.

Online inference

Chalk’s online store & feature computation engine make it easy to query features with ultra low-latency for your online inference use cases.

Integrating Chalk with your production application takes minutes via Chalk’s simple REST API:

> chalk query \
      --in \
      --out user.identity.is_voip_phone \
      --out user.fraud_score \
      --staleness user.account_balance=10m \
      --environment staging \
      --tag live

Features computed to serve online requests are also replicated to Chalk’s offline store for historical performance tracking and training set generation.

Offline training

Data scientists can use Chalk’s Jupyter integration to create datasets and train models. Datasets are stored and tracked so that they can be re-used by other modelers. Chalk implements model provenance to track inputs, outputs, and other data for audit and reproducibility.

Chalk AI - Documentation Reference
Jupyter Notebook
Chalk AI - Alerts

Chalk datasets are always “temporally consistent.” This means that you can provide labels with different past timestamps to easily get historical features that represent what your application would have retrieved online at those past times. Temporal consistency ensures that your model training doesn’t mix “future” and “past” data.

label 1
label 2
Feature 1
Feature 2
Feature 3