If you like what we're working on, please  star us on GitHub. This enables us to continue to give back to the community.

What to Look for When Monitoring for Performance Analysis

If you would like to contribute your own blog post, feel free to reach out to us via blog@deepchecks.com. We typically pay a symbolic fee for content that’s accepted by our reviewers.


If you have any experience with Continuous Integration/Continuous Deployment then you know how much of a hassle it is. The configuration files themselves will boggle your mind and then integrating the model to cloud service is another challenge. It is a pretty hefty task. And once you deploy the model you might think that the job’s done but it isn’t. The moment you deploy the model it starts to degrade over time as the model is dynamic and sensitive to the changes that happen to the data in the real world.

You can assume that the model naturally overfits once you deploy it. This is why deployment should not be your final step.

Performance degradation is one of the limitations of ML models. It must be nurtured or retrained with new data frequently so that the performance is consistent and robust at all times. With that being said, our goal is to maintain the model’s performance as long as it is operational.

Monitoring Performance Analysis

We scrutinize the following areas when monitoring the performance of the model. Performance analysis is an important part of the MLOps monitoring process, as it can help to ensure that the ML models are effective and deliver the desired results.


Data is the fuel of an ML model. Its quality and integrity are essential in ensuring that machine learning models are trained on high-quality data so they yield accurate and reliable results.

One of the major issues with data is data drift. Data drift refers to a phenomenon in which the fundamental quality of the data changes over time, leading to a degradation in the model’s performance.

Below we will discuss some of the key points that will enable you to monitor and delay data drift as much as possible, since it is impossible to maintain a consistent distribution of data. As human behavior changes over time, so  does data distribution. 

Testing. CI/CD. Monitoring.

Because ML systems are more fragile than you think. All based on our open-source core.

Deepchecks HubOur GithubOpen Source

Covariate/Feature Drift

Covariate drift, also known as feature drift, occurs when the statistical properties of the features (or covariates) in a dataset change over time.

Label Drift

Label drift generally occurs when the distribution of the target variable in a dataset changes over time.

Concept Drift

This happens when the underlying relationship between the features and the target variable in a dataset changes over time.

Prior Probability Shift

A prior probability shift takes place in a dataset when the prior probabilities of the different classes in a dataset change over time.

Other data factors to look for while monitoring for performance analysis are:

  1. Presence of outliers. These are data points that are significantly different from the majority of the data. These data points can have a significant impact on the performance of machine learning algorithms, as they can distort the overall distribution of the data and can cause the algorithms to make incorrect predictions.
  2. Data integrity. It refers to the accuracy, completeness, consistency, and reliability of data. It is essential for ensuring the accuracy and usefulness of data.
  3. Data quality. It measures how well the data meet the requirements and expectations for the purpose of the ML task. It is an essential part of developing any machine learning system.
  4. Preprocessing pipeline. This is a series of steps or transformations applied to raw data to prepare it for use in a machine learning model. Preprocessing pipeline provides consistency, uniformity, and usability, which helps to improve the performance and reliability of the ML models

Machine Learning Model

The capacity of the model is the reflection of the complexity of the data itself. An ML model can be considered a good model if its consistency remains the same. One way to maintain consistency is to check for model drift.

Another important concept is Model Drift or Concept Drift. This is different from data drift. Model drift refers to a change in the performance or predicting capability of a machine learning model over time. This is due to changes in the data such as changes in the underlying distribution of the data, changes in the data collection process, or changes in the business context.

To analyze model drift, the first step is to define baseline performance metrics of the ML model, such as accuracy or precision, and track these metrics over time. This can help to identify any changes in the performance of the model that may be indicative of model drift.

Other important factors that can be associated with model performance are:

  1. Accuracy can be determined using metrics such as precision, recall, and F1 score.
  2. Speed refers to the amount of time it takes for the model to process data and make predictions.
  3. Robustness is the ability of the ML model to continue making accurate predictions even when faced with noisy, incomplete, or unexpected data. Robust models can be more useful in dynamic environments where the data may be constantly changing for instance radiology.
  4. Consistency in models lowers the likelihood of sudden drops in performance, which can be problematic in applications where the data is evolving rapidly.
  5. Explainability is very useful in applications where the reasons for the predictions are important (e.g., medical diagnoses, credit decisions).
  6. Bias is the presence of systematic errors or preferences in the predictions made by the ML model. Models with high levels of bias can be unfair or discriminatory and should be avoided in most applications.

Context Monitoring

Context refers to the information or background knowledge that is available to the model and can be used to help it make more accurate predictions. This can include information about the specific task the model is being used for, the business model, the targeted audience, the geographical region of interest, the data it is processing, or the environment in which it is operating.

Context monitoring ensures that the correct data is being collected and preprocessed in a specific way according to the requirement.

When it comes to monitoring an ML model, engineers should keep in mind that the context or business goal of the model is not lost. Context monitoring allows engineers to keep the entire pipeline of the system aligned with purpose.

Infrastructural Monitoring

In MLOps, infrastructural monitoring refers to the process of monitoring the underlying infrastructure that is used to train and deploy machine learning models. This can include monitoring the performance of the hardware and software used to run the ML models, as well as monitoring the availability and reliability of the infrastructure–cloud services, containers, and scalability.


Monitoring is an important phase in MLOps as it helps to ensure that the machine learning models are performing as expected and delivering the desired results. By regularly monitoring the performance of the ML models and the behavior of the data, MLOps teams can identify any issues or anomalies that may be impacting the model’s accuracy or reliability.

Testing. CI/CD. Monitoring.

Because ML systems are more fragile than you think. All based on our open-source core.

Deepchecks Hub Our GithubOpen Source

Recent Blog Posts

Reducing Bias and Ensuring Fairness in Machine Learning
Reducing Bias and Ensuring Fairness in Machine Learning

Testing your NLP Models:
Hands-On Tutorial
March 29th, 2023    18:00 PM IDT

Register Now