Kubeflow

Helps orchestrate machine learning training pipelines across on-prem and cloud-based resources.
Can containerize training and serving infrastructure.

Components

Support for distributed TensorFlow training via the TFJob CRD
- TFJob is a Kubernetes custom resource used to run TensorFlow training jobs on Kubernetes.
The ability to serve trained models using TensorFlow Serving
- Flexible, high performance serving system for machine learning models.
- Installed by default when deploying Kubeflow.
A JupyterHub installation with many commonly required libraries and widgets included in the notebook installation, included those needed for TensorFlow Model Analysis (TFMA) and TensorFlow Transform (TFT)
- JupyterHub – Multi user server for Jupyter notebooks
- TFMA – Library for evaluating TF models.
  - Uses Apache Beam
- TFT – Preprocessing/feature engineering
  - Can prevent Training Serving Skew
    - Difference between performance during training and performance during serving. This can be caused by:
      - Discrepancy between how you handle data in the training and serving pipelines.
      - Change in the data between when you train and when you serve.
      - Feedback loop between model and algorithm.
  - Uses Apache Beam
Kubeflow Pipelines
- Has UI for managing and tracking experiments, jobs, and runs.
- An engine for scheduling multi-step ML workflows
- An SDK for defining and manipulating pipelines and components
- Notebooks for interacting with the system using the SDK