The ML Pipeline in MLOps

Every real-world ML system is made of 3 programs (aka pipelines): → Feature pipeline → Training pipeline → Inference pipeline

Every real-world ML system is made of 3 programs (aka pipelines):

→ Feature pipeline
→ Training pipeline
→ Inference pipeline

And this is how they work ↓

1️⃣ 𝗧𝗵𝗲 𝗳𝗲𝗮𝘁𝘂𝗿𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 fetches raw data, from
\

• a data warehouse\
• an external API, or\
• a website, through scrapping
  
  and generate features, aka the inputs for your ML model, and stores them in a Feature Store so that the other 2 pipelines can later use these features.
  
  2️⃣ 𝗧𝗵𝗲 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 takes the features from the store and outputs a trained ML model.
  
  These are (in general) the best models for each domain:
  \
• Tabular data → XGBoost\
• Computer Vision → Fine-tune a Convolutional Neural Net\
• NLP → Fine-tune a Transformer net.
  
  3️⃣ 𝗧𝗵𝗲 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 takes recent features and a trained model, and outputs predictions that are consumed by a downstream service or client.
  
  This pipeline can run\
• on a schedule -> batch-scoring system\
• on-demand -> online prediction, like a REST API or Stream app.