Feature Stores

Misc

  • Benefits

    • Increases Reproducibility
      • Easily track the version of the features used in each model and reproduce the model’s results if needed.
      • Useful in a collaborative environment where multiple people are working on the same project.
    • Discovery and Testing Features Easier
      • Having features in a centralized location makes comparing the performance of various features and versions of those features easier.
    • Easier to Scale
      • It’s easier to share features between ML models which means fewer resources (e.g. development, deployment) will be required. Allowing more models to be added more efficiently and cheaply.

  • Features of a Feature Store

    • Designed with ML modelling in mind
      • Can handle large amounts of data and perform feature engineering at scale
    • Handles versioning of features
      • Easy to track which features were used for a particular model, making it simple to reproduce or deploy the model in the future
    • Allows for different levels of access control
      • A data scientist can work on a feature without worrying about affecting other users which can’t be said about warehouses

  • Best used if you have a substantial number of features that are computationally expensive, frequently improved, and used in many ML models.

    • Here “Transform” is referring to something like an AWS lambda function that’s triggered to transform the data

    • Cases where adding a feature store adds unnecessary complexity:

      • Feature value needs to “seen” by the client (e.g. app)
        • Not exactly sure why this matters or what “seen” means
        • Maybe this is a latency thing?
      • Feature is in a data warehouse.
      • Feature isn’t time dependent
        • So only streaming and not batch serving I think
      • Computationally inexpensive

    • Example

      • Embedding of a song, artist, and user features in a music streaming service.
        • There is a team updating user and song embeddings on a daily basis. Every time the model that consumes this feature, it is retrained — high commercial value use cases will need to re-train periodically — the training code will need to fetch the values of this feature that align with the training labels and the latest version of the embedding algorithm.

  • Positioning within a pipeline

    • Looks like something dbt (“Feature Pipelines”) would write to.
    • Reminds me of the description of a data mart.

  • Connectors

    • TensorFlow’s TFXI (TensorFlow Extended Input/Output)
      • Module allows you to easily read data from Feature store and feed it into your TensorFlow model.
      • Supports data preprocessing, so you can do things like normalization and feature selection right within TensorFlow.
    • PyTorch’s DataLoader
      • Class that allows you to easily read data from Feature store, process, and feed it into your PyTorch model.

Brands

  • Google Vertex AI feature store
  • Amazon SageMaker Feature Store
  • Databricks Feature Store
  • HopsWorks Feature Store
  • tecton.ai
    • site
    • cloud agnostic
  • bytehub
  • Feast
    • It is a standalone, open-source feature store that organizations use to store and serve features consistently for offline training and online inference.
  • DataRobot
  • Algoworks
  • Hugging Face: A feature store for natural language processing (NLP) models that allows for easy sharing and management of pre-trained models and features.