Skip to content

Deployment

Our deployment model is designed to keep your data safe and isolated as much as possible in your cloud, while keeping it easy to integrate with the rest of your stack. In this section, we will explain Tecton's deployment model, and the various compute engines, storage systems, and data sources we support.

Deployment Model

The SaaS Deployment model divides responsibility between a control plane and a data plane.

The data plane, which lives entirely in your AWS account, is responsible for all data processing andĀ data storage. It connects to your raw data sources, turns raw data into features, and stores those features in the offline and online feature store. The data plane hosts only AWS native services.

The control plane, which lives in Tecton's AWS account, is operated by Tecton and manages Tecton's metadata as well as core services (feature serving, pipeline orchestration, monitoring, web-ui). The control plane communicates with the data plane to orchestrate data pipelines, monitor its health, and serve features.

SaaS Deployment Model

This SaaS architecture minimizes Tecton's footprint in your AWS account and is limited to the processing and storage of sensitive data using exclusively AWS native services. Read more about this architecture's benefitsĀ here. If you have needs beyond this SaaS architecture, contact us and let's discuss.

Note

Please note that these options are currently AWS-based. At this time, Tecton does not support currently Azure or GCP. If your requirements are not supported or want to learn about other possible deployments, please contact our sales team to set up a discussion.

Compute Engine for Feature Processing

Apache Spark is currently Tecton's primary compute provider for running all feature transformations. Tecton orchestrates jobs using your existing Databricks or AWS EMR deployments. More compute providers, such as popular data warehouses, are coming soon.

Storage

Tecton use different storage strategies for online vs. offline feature storage.

  • To enable low-latency online access, feature values are written to DynamoDB. More key-value store options are coming soon
  • To enable offline access to historical data for training, feature values into Parquet files. These are stored in S3 buckets in your AWS account

Data Sources

Tecton currently supports the following data batch data sources:

  • CSV, Parquet, and JSON Data Sources on S3
  • Hive Tables via AWS Glue Data Catalog
  • AWS Redshift Tables
  • Snowflake Tables

Tecton currently supports the following stream data sources:

  • AWS Kinesis Streams
  • Kafka Topics

The Tecton team continually adds support for new data sources. If you need to use a data source that isn't listed above, reach out to our sales team.

Please contact our sales team if you have additional needs from what's listed here.

Service Level Objectives

For valid requests, Tecton provides 99.95% availability and guarantees 99% will complete in 100ms or less. See additional details and validity requirements here.

Monitoring

Tecton helps monitor both data quality and operational performance, so your models can reliably receive fresh feature data within latency requirements.

Currently, Tecton monitors the following operational metrics:

Tecton will also soon provide data quality monitoring and data drift monitoring. Stay tuned!

Next Steps

Next we will explore Tecton's framework in depth, where we will go through how each entity in the framework works, starting with the Feature View..

If you don't already have access to a Tecton cluster, please contact us and we can show you how Tecton can power your models and drive results.