Infrastructure & Tooling for MLOps

Storage & Compute Layer

• all the compute ressources a company has access to and the mechanisms to determine how theses ressources can be used
• can be used into smaller compute units to be used concurrently threads : to execute a job instance : “permanent” unit

  Dev Environment

  IDE, versionning, CI/CD, SSH, containers

  Ressource Managment

• CRON : run script at a predetermined time & tells if failed or succeded
• SCHEDULER : cron that handle dependencies (needs to know the available ressources)
• ORCHESTRATORS

  DS Workflow managment

  specify workflows as DAGs where each step is an edge Airflow, Argo, Perfect, Kubeflow, Metaflow

  ML Plateform

  shared set of tools for ML Deployment

• Model Store :
  • model definition
  • model parameters
  • featurize & predict functions
  • dependencies
  • data (version or endpoint)
  • model generation point
  • experiment artifacts
  • tags ML Flow is the most popular
• Feature Store
  • feature managment
  • feature consistency
  • feature computation