Anton Čorňák
0 %
Your Photo
Software Engineer

0 %
0 %
0 %

React0 %
Next.js0 %
Python0 %
Golang0 %
TypeScript0 %
JavaScript0 %
Docker0 %
Google Cloud Platform0 %
Kubernetes0 %
Amazon Web Services0 %

Git, GitHub
Agile, Jira, Confluence

MLOps Guidelines pt. 1: Your model is ready, your journey is just beginning
September 01, 2023
MLOps Guidelines pt. 1: Your model is ready, your journey is just beginning

You're a junior MLOps engineer at a mid-sized tech company. You've been tasked with developing a machine learning model to improve the recommendation system for your company's e-commerce platform. After weeks of feature engineering, hyperparameter tuning, and cross-validation, you have it - your model boasts a 95% accuracy rate - and that alone is an achievement worth celebration. You're eager to show off your results, but then the question hits you: "What now?"

The model is sitting on your local machine, and you realise that turning this mathematical marvel into a functional part of your company's tech stack is a whole new challenge. How do you take it from Jupyter Notebook to your company's servers? How do you ensure it performs as expected with new data? How do you update it? If these questions leave you scratching your head, you're not alone. And that's where MLOps comes in.

Understanding MLOps

What MLOps Means

MLOps, or Machine Learning Operations, is the practice of combining machine learning engineering with the principles of DevOps. The goal is to streamline the end-to-end ML lifecycle, from data collection to model training, deployment, and monitoring. MLOps ensures that machine learning models are not just a one-off project but an integrated part of a continuously evolving system.

How it Connects to DevOps

DevOps (Development and Operations) emphasises the automation and monitoring of software development, from integration to delivery and maintenance. MLOps extends this philosophy to machine learning models. Just as DevOps ensures seamless integration between software development and operations, MLOps seeks to facilitate the collaboration between Data Scientists, ML Engineers, and Operations teams. It provides the tools and practices needed to deploy machine learning models into production reliably and efficiently.

The Significance of MLOps in a Machine Learning Project

A machine learning model is only as good as its implementation and upkeep. A trained model on a laptop is a mathematical curiosity; a model in production is a business asset. MLOps bridges this gap. It addresses the "last mile" issues that come with transitioning from a training environment to a real-world setting. Challenges like model validation, scalability, and monitoring are addressed under the umbrella of MLOps.

Understanding MLOps is crucial for anyone tasked with turning machine learning models into production assets. With the knowledge and guidelines provided in this series, you'll be better prepared to tackle the next steps of your machine learning journey.

Pre-Deployment Checklist

Before your model can be unleashed into the wild, it needs to go through a series of checks and preparations. Here's a list of crucial steps to ensure your model is production-ready.

Model Validation and Testing

Once a model is trained, validating its performance is the first step towards deployment. Cross-validation, backtesting, and A/B tests are some methods used to ascertain how the model performs on unseen data. These tests provide insights into how your model will behave in a production environment and may also reveal any potential flaws or biases. We will focus on this part later in this guidelines series.

Data Versioning

Just like you maintain versions of your codebase, maintaining versions of your dataset is equally important. Data versioning ensures reproducibility and traceability in your machine learning workflow. Whenever the data changes, a new version can be created, making it easier to roll back to previous states or experiment with different data configurations.


Documenting your work might not seem like the most exciting task, but it is crucial for the longevity and maintainability of your project. A well-documented model will include explanations of the model architecture, feature importance, and any tuning parameters. Moreover, operational documentation can help other team members understand how to deploy and maintain the model, or what steps to take if something goes wrong.

These are essential steps in preparing your model for deployment. Proper attention to validation, data versioning, and documentation not only streamlines the transition to production but also sets the stage for long-term success.

Deployment Options

Once your model is validated, versioned, and well-documented, the next logical step is deployment. Here are some options you can consider, each with its own set of advantages and challenges.


Deploying your model on-premises means that you're using your own hardware infrastructure. This approach offers full control over the data, which can be a significant benefit for businesses with sensitive or proprietary information. However, this comes with the responsibility of managing your own servers, scaling resources, and ensuring uptime, which can be resource-intensive.

Cloud-based Services

Cloud-based options, like AWS SageMaker, Google Cloud ML, or Azure ML, offer a range of services that handle much of the heavy lifting involved in deployment. These platforms provide scalability and are generally more flexible than on-premises solutions. They can automatically adjust to different loads and can be more cost-effective due to their pay-as-you-go model. However, you do give up some level of control, and there are considerations around data security and compliance.


A hybrid solution combines both on-premises and cloud-based resources. This might be suitable for organisations that have specific data that must remain on-site but still want to take advantage of the scalability and tools that cloud services offer. It allows for a balanced approach, giving you the best of both worlds, though it also involves managing two different environments.

Choosing the right deployment option should be based on your project's specific needs, taking into consideration factors like data sensitivity, scalability, and available resources.

Continuous Integration and Continuous Deployment (CI/CD)

Moving beyond initial deployment, one must consider the long-term maintenance and improvement of machine learning models, which brings us to the concept of Continuous Integration and Continuous Deployment (CI/CD).

Importance of CI/CD in MLOps

CI/CD automates the steps that come after writing code, reducing manual effort and the possibility of errors. This is especially significant in MLOps, where models require constant updates to reflect new data or insights. CI/CD ensures that your machine learning models are as dynamic as the data they are based on, thereby increasing efficiency and reducing time-to-market for improvements.

Basic CI/CD Pipeline for Machine Learning

A typical machine learning CI/CD pipeline consists of several stages:

  1. 1.  Code Repository: All machine learning code is stored in a version-controlled repository.
  2. 2.  Automated Testing: As new code is committed, automated tests are run to check for errors and ensure that the model is producing the expected outputs.
  3. 3.  Model Training: After passing tests, the model is re-trained with the latest data.
  4. 4.  Model Validation: Post-training, the model is again validated to ensure that it meets performance benchmarks.
  5. 5.  Deployment: If all checks pass, the updated model is deployed to replace the older version.
  6. 6.  Monitoring: Finally, the model's performance is continuously monitored, and any deviations trigger alerts for further investigation.

Implementing a CI/CD pipeline streamlines the maintenance and updating of machine learning models, making it easier to sustain a project long-term.

Monitoring and Maintenance

After deployment, the work is far from over. A model's performance can deteriorate over time, or unforeseen issues could arise. This makes ongoing monitoring and maintenance critical components of MLOps.

Model Drift

Model drift occurs when the statistical properties of the target variable, which the model is trying to predict, change over time. This inevitably degrades the model's performance. Monitoring for model drift and retraining your model accordingly is crucial for maintaining its effectiveness.

Data Quality Checks

Poor quality or inconsistent data can wreak havoc on your model's performance. Implementing automated data quality checks as part of your monitoring system can catch issues like missing values, outliers, or incorrect labels before they impact your model.

Performance Metrics

Continuous monitoring isn't just about watching for negative changes; it's also about understanding how your model is performing in general. Metrics like accuracy, precision, recall, and F1-score can give you insights into your model's performance. More domain-specific metrics may also be relevant depending on the application.

By setting up comprehensive monitoring and conducting regular maintenance, you can ensure that your model remains robust and effective in a production environment.


Navigating the MLOps landscape can seem overwhelming, especially once a model has been trained and you're wondering what comes next. To recap, the journey post-training involves understanding MLOps, setting up pre-deployment checks like model validation and data versioning, selecting an appropriate deployment option, and finally, setting up a CI/CD pipeline. All of this should be underpinned by a solid monitoring and maintenance strategy to ensure long-term performance.

In upcoming posts in the MLOps Guidelines series, we'll delve deeper into each of these areas, providing more detailed guides and best practices to follow.

What's Next in MLOps Guidelines

The next part of the series will focus on MLOps terminology and key concepts. We'll break down commonly used terms, acronyms, and foundational principles essential to understanding the MLOps landscape that are often assumed to be understood, but rarely explained. Consider it your MLOps 101, serving (pun intended) as a comprehensive glossary that will make subsequent topics more accessible and meaningful.

Stay tuned!

Posted in



Anton Čorňák





Liked the post? Leave a feedback!
© 2024 All Rights Reserved.