Tecton

Tecton is a feature store enables data scientists to build a library of great features and serve them in production.

Combine Batch, Streaming, and Real-Time Data

Use all your enterprise data to build high-quality features. Combine batch data (e.g. Amazon S3, Amazon Redshift, Snowflake), streaming data (e.g. Apache Kafka, Amazon Kinesis), and real-time data. Real-time data is passed to Tecton at the time of the feature request, and Tecton executes on-demand transformation to generate real-time values.

Manage features as code

Define features in Python files that contain the transformation logic and metadata needed to compute features on an ongoing basis. Version-control your features (e.g. in Git) and integrate them with your existing Code Review and CI/CD processes.

Develop with familiar data science tools

Build features using familiar programming languages and libraries including Python, SQL and PySpark. Use Tecton’s Python SDK in your preferred notebook environment to create training datasets.

Generate accurate training data

Create accurate training datasets with just a few lines of code. Use row-level time travel to deliver the right values at the right time, for each individual row.

Deploy

Plan feature changes with confidence

“Am I about to modify a feature used in production? How much will this feature cost to process? Is this new feature a duplicate?” Before applying your changes, Tecton generates a plan that allows you to answer these questions. Test changes in private workspaces before deploying them to your production environment, and integrate seamlessly with your existing CI/CD pipeline.

Automate feature transformations

Tecton orchestrates data pipelines to generate backfills and continuously compute fresh feature values. Alternatively, ingest feature data from pipelines managed outside of Tecton.

Ensure data consistency

Provide a single source of truth for feature data across your organization. Tecton stores historical data in offline storage, and fresh data in online storage for low-latency retrieval. Ensure data consistency over time and eliminate training / serving skew. Reproduce historical data sets with row-level time travel to deliver the right values at the right time, for each individual row.

Serve features online

Retrieve feature values from Tecton’s REST/gRPC API to power real-time predictions in production at high scale. Median serving latencies are ~5ms and can easily scale to 10,000s of requests per second.

Monitor data quality and operational metrics

Monitor features at every step of their lifecycle. Validate ingested data and detect data drift. Monitor operational metrics including serving latencies, serving volumes, and storage consumption.

Share

Search and discover features

Build a unified catalog of features across your organization. Enable teams to search and discover existing features to enable re-use across teams and models. Track detailed information on each feature including data lineage, feature health, and feature value distribution to increase confidence in existing features

Enterprise-Ready

Enterprise-grade security

Tecton is deployed in your VPC, so your data never leaves your cloud account. Control access to individual features and data with Role-Based Access Control and ACLs. Integrate with AWS IAM, Okta, or any solution that supports SAML or Oauth.

Built for scale

Based on Uber Michelangelo’s battle-tested architecture blueprint, Tecton is built for scale and can support thousands of models, tens of thousands of features, and millions of predictions per second. You can start small, and Tecton will scale with you.