sennder MLOps Primer

About the author: Varun Chitale is an experienced Machine Learning Engineer Manager with more than 6 years of experience. He has worked in ML applications of all kind with particular focus on NLP and Recommendation Systems.

How We Explored And Selected MLOps Technologies

At Sennder, we're proud to be at the forefront of machine learning and AI technology. Our team of skilled data scientists and engineers work tirelessly to develop cutting-edge models that deliver value to our customers and partners.

But as any seasoned data science practitioner knows, building a machine learning model is only half the battle. The real challenge lies in productionizing it - in deploying it at scale, ensuring it's performant, and keeping it up to date as data and business requirements evolve.

That's why at Sennder, we've invested heavily in the development of a robust machine learning engineering platform. Our platform includes several key components that are critical to productionizing an ML model, including an ML pipeline orchestrator, ML metadata and artefact tracking, a model registry, and tools for model training and serving.

To make sure we chose the right tools for each of these components, we used the process of Markdown Any Decision Records (MADR). This helped us make informed decisions that were well-documented and transparent.

MADR (Markdown Any Decision Records) is a process for documenting decisions made during a project or initiative. It involves creating a simple Markdown file that outlines the decision, its rationale, and any relevant context or background. MADR files can be easily shared and reviewed, making it easier for team members to stay informed and make informed decisions going forward. Overall, MADR helps teams make better decisions and promotes transparency and collaboration.

In this blog post, we'll dive deeper into each of these components, explore the decision-making process behind our tooling choices, and showcase some of the amazing work our team has done with this technology. We hope this post inspires other data science practitioners and potential team members to join us in our mission to push the boundaries of what's possible with machine learning.

The Components

ML Pipelines Orchestrator

An ML pipelines orchestrator is a tool or platform that manages the flow of data and computations during the machine learning process. Essentially, it helps you automate the steps involved in training and deploying an ML model, so that you can more easily manage the complexity of these processes. A common example of an ML pipelines orchestrator is Apache Airflow, which allows you to create DAGs (Directed Acyclic Graphs) to define the sequence of tasks that need to be executed.

Thanks to our use of MADR, we were able to carefully evaluate and document our decisions about which ML Pipelines Orchestrator tools to consider, making it possible to compile this comprehensive list for your reference.

Model Training

Model training is the process of training your machine learning models on data to improve their performance. This involves selecting the appropriate algorithms and hyperparameters, preparing the data, and running the training process itself. A common example of a tool for model training is scikit-learn, which provides a wide range of algorithms and tools for preprocessing data and running the training process.

Table 2: Platforms For Distributed Training

ML Metadata And Artefacts Tracking

ML metadata and artefacts tracking is the process of capturing and storing information about the data and models used in the machine learning process. This helps ensure reproducibility and auditability, as well as facilitating collaboration and experimentation. A common example of an ML metadata and artefacts tracking tool is MLflow, which allows you to log information about the data, models, and experiments you run during the machine learning process.

You know the drill, here’s the compiled data for different tools for this task –

Model Registry

A model registry is a tool or platform that allows you to store and manage different versions of your machine learning models. This helps you keep track of the changes made to your models over time, and makes it easier to deploy and manage them in production. A common example of a model registry is the TensorFlow Model Garden, which provides a centralized repository of pre-trained models and tools for managing model versions and deployments.

Drill –

Model Serving

Model serving is the process of making your machine learning models available to users or applications for inference. This involves deploying the model to a production environment and providing an API or other interface for accessing it. A common example of a model serving tool is TensorFlow Serving, which allows you to deploy TensorFlow models to production environments and provides an API for making predictions.

The Final Stack

Together, that is a lot of information to digest. Individually, each of these were parallelised for simultaneous completion.

Next, we used the popular MoSCoW Method to evaluate the tools and decide the winners. Here's a crisp explanation of MOSCOW prioritization:

• MOSCOW prioritization is a project management technique to prioritize requirements or features based on their importance.
• It stands for Must have, Should have, Could have, and Won't have.
• Must-have requirements are critical and necessary for the project's success.
• Should-have requirements are important but not critical.
• Could-have requirements are desirable but not essential.
• Won't-have requirements are excluded from the project scope.
• MOSCOW prioritization helps teams focus on the most important features and manage stakeholder expectations.
• It also helps reduce scope creep and ensure the project delivers the most value with available resources.

Azure was a close runner-up in some cases, however, Sagemaker beat it because of it’s wide applicability.

In conclusion, choosing the right frameworks for building a robust and scalable machine learning architecture can be a challenging task. However, the MADR approach can be a valuable tool in streamlining the decision-making process. By evaluating various tools and documenting the decision-making criteria and rationale, teams can ensure transparency and accountability in the decision-making process. While extensive research may seem time-consuming, it is essential to make an informed decision that aligns with the organization's goals and requirements. In the end, a well-designed machine learning architecture can significantly impact the success of ML projects, and investing in the right tools and frameworks can make all the difference.