TodayTools

MLOps flashcards

Take machine learning models from a notebook to reliable production: tracking, versioning, pipelines, serving, monitoring and LLM operations.

Donate
Beginner friendly, starts from zero
Section
How to study
0
MLOpsStart from zero · 1 of 10

What is MLOps, in one plain sentence?

MLOps is the practice of building, releasing and looking after machine learning so they keep giving good answers in real use, not just once in a notebook.

ML plus operations: build it, then keep it healthy.
Show again
10 left
Got it

Tap to turn over. Swipe right if you know it, left to see it again.

All topics

What you will learn

  1. Start from zero

    • What MLOps is4 cards
    • The key words4 cards
    • Getting started2 cards

  2. Core basics

    • The lifecycle3 cards
    • Experiment tracking3 cards
    • Reproducibility3 cards
    • Packaging and registry3 cards
    • First deployment4 cards

  3. Next steps

    • Pipelines and automation4 cards
    • CI/CD for models3 cards
    • Safe rollouts3 cards
    • Monitoring4 cards
    • Documentation2 cards

  4. Go deeper

    • Serving at scale4 cards
    • LLM operations4 cards
    • Failure modes4 cards
    • Security and governance4 cards

Final test and certificate

10 questions drawn fresh from the whole topic, about 15 minutes. Sign in, score 80% or more and get a free certificate with your name and a public check link anyone can open, such as an employer or a teacher. Not signed in? Try a practice test from the cards.

Take the final test

About these cards

  • Reviewed 26 October 2026
  • CoversMLflow 3.17Python 3.12+
How these cards are made

MLOps at a glance

MLOps is the set of habits and tools that keep machine learning models working reliably after they leave a data scientist's laptop.

Training a model is like perfecting a recipe at home. MLOps is running the restaurant kitchen: same taste every day, fresh ingredients, checks on every plate, and a quick fix when something goes wrong.

Words you will meet

Model
A program that learned patterns from examples and makes predictions
Training
Showing the model many examples so it learns the patterns
Experiment
One training try with certain settings, whose results you record
Deployment
Putting a model where real apps and people can use it
Pipeline
A chain of automatic steps, like get data, train, test, deploy
Version
A saved, numbered copy of data, code or a model you can go back to
Monitoring
Watching a live model's inputs, outputs and accuracy over time
Drift
When real-world data slowly stops looking like the training data

The steps

  1. 1. Start from zero (10 cards): Why models need care after training, the life of a model, and the plain words used later.
  2. 2. Core basics (16 cards): The model lifecycle, tracking experiments, reproducibility, packaging and first deployment.
  3. 3. Next steps (16 cards): Pipelines, CI/CD for models, safe rollouts, monitoring and retraining: the weekly work of an ML engineer.
  4. 4. Go deeper (16 cards): Scale, cost, LLM operations, security, governance and failure modes: what a senior ML engineer handles.
All 10 cards in Start from zero
What is MLOps, in one plain sentence?
MLOps is the practice of building, releasing and looking after machine learning models so they keep giving good answers in real use, not just once in a notebook.
Why do models need looking after once they work?
The world changes. Prices, habits and seasons shift, so a model trained on last year's data slowly gets worse. Data pipelines also break, and nobody notices without checks.
Where do you meet models that need MLOps?
Card fraud alerts, film and shopping recommendations, delivery time estimates, spam filters, loan decisions and chat assistants all run models that are retrained and watched.
What everyday thing is MLOps like?
Running a restaurant kitchen instead of cooking once at home: written recipes, fresh ingredients, the same taste every day, checks on every plate, and quick fixes when a dish goes wrong.
What are a model, training and an experiment?
A model is a program that learned patterns from examples. Training is the learning step. An experiment is one training try with chosen settings, whose results you write down to compare.
What are deployment and a pipeline?
Deployment means putting the model where apps can use it, like behind a web address. A pipeline is a chain of automatic steps: get data, clean, train, test, deploy.
What are versions, and why version data and models too?
A version is a saved, numbered copy you can return to. If you version only code, you cannot rebuild last month's model, because the data and settings have changed.
What are monitoring and drift?
Monitoring watches a live model's inputs, outputs and accuracy. Drift is when real data slowly stops looking like training data, for example a new payment method appears.
What do you need before learning MLOps?
Basic Python, how a machine learning model is trained and tested, and git for versioning code. Docker basics help in the later levels.
What will you be able to do after this topic?
Track experiments, version data and models, package and serve a model, automate training with pipelines and CI/CD, monitor drift and run LLM apps with proper checks.
MLOps cheatsheet

MLflow

  • mlflow.set_experiment("churn") Group runs
  • mlflow.autolog() Log automatically
  • mlflow.log_param / log_metric Record settings and scores
  • mlflow.sklearn.log_model(m, name="m") Save a LoggedModel
  • client.set_registered_model_alias(n, "champion", v) Promote a version
  • mlflow.pyfunc.load_model("models:/n@champion") Load by alias
  • @mlflow.trace Trace an LLM step

Data and environment

  • dvc add data/x.parquet Track a dataset
  • dvc push / dvc pull Sync data with remote
  • uv lock && uv sync --frozen Pinned environment
  • docker build -t model:1.4.0 . Versioned image
  • mlflow server --backend-store-uri Run a tracking server

Release patterns

  • Shadow Live traffic, answers unused
  • Canary Small share first
  • Blue-green Two full stacks, switch
  • A/B test Random split, business metric
  • Rollback Move alias or traffic back

What to monitor

  • Latency p50 / p99, errors Service health
  • Schema and ranges Input quality
  • PSI, KS test Data drift
  • Prediction distribution Early warning
  • Accuracy on fresh labels Real quality
  • Cost per 1,000 predictions Spend
Donate