Introduction to MLOps on Vertex AI

1. Introduction to MLOps on Vertex AI

Now that you’re familiar with what Vertex AI is, let’s explore how Vertex AI can help with MLOps processes. If you worked with ML models before, you know that training and deploying ML models can be time-consuming, because you need to repeatedly add new data and features, try different models, and tune parameters to achieve the best result. To solve this problem, organizations need to build the necessary ML engineering culture and capability. If we consider that MLOps is a set of standardized processes and technology capabilities for building, deploying, and operationalizing ML systems rapidly and reliably, it can be the answer that you’re looking for in order to build ML engineering culture and capability. Before we look at workflows and capabilities needed in each phase of the MLOps lifecycle, let’s explore the MLOps concept in depth: MLOps aims to unify ML system development or ML with ML system operations or Ops. It advocates automating and monitoring every step of ML system construction. And finally, it maintains aligned versions of data and models alongside code and components. At this point, it is fair to ask this question: How does MLOps achieve each of these goals? To answer this question, it’s important to understand what ML engineering is and how ML engineering is tightly coupled with data engineering and app engineering: Let’s start with ML engineering. ML engineering is a superset of the discipline of software engineering and is designed to handle the unique complexities of the practical application of ML. MLOps is one of the methodologies adopted by ML engineering that focuses on operationalization and automation of model development, training, deployment, and governance processes. To learn more about ML engineering, check the reading at the end of this section. Let’s continue with data engineering. Data engineering focuses on ingesting, managing, and processing data to prepare datasets and features. All of these outputs created by data engineering, datasets and features, can be used in enterprise data warehouses, business intelligence or BI systems and ML engineering. And finally, there is app engineering. App engineering focuses on designing, developing, or migrating applications. Deploying ML models and integrating them with your applications and systems is the main subject of app engineering when you build an ML application. That means that engineering an ML enabled solution is an iterative process between the three disciplines: ML engineering uses the datasets and features produced by the Data Engineering team and feeds back the data requirements to the Data Engineering team. ML engineering also produces the models deployed by the App Engineering team, which in turn shares the application’s key performance indicators, or KPIs, with the ML Engineering team. Therefore, ML engineering should not be performed in isolation. Instead, it should use existing investments in DataOps and DevOps and integrate with the data engineering and app engineering to be effective. By the inspiration of this integration, the MLOps lifecycle can be divided in six iterative processes, and two cross-cutting processes in the center: The lifecycle starts with the ML development process, which is the developing and prototyping process of your ML models. This is mostly the experimentation phase of ML model development. The second process in the lifecycle is training operationalization, which refers to checking whether your model works on the production. This includes testing internal and external data connections and configurations and putting the model routine and a stable operational phase after the first experimentation phase is performed. Then, in the continuous training process, you retrain your production models with the new data. Next is the model deployment process, where actual continuous integration and delivery of your models occurs. The prediction or inference serving process refers to hosting ML models as services to serve online predictions or as part of a batch prediction system to serve offline predictions. And continuous monitoring is the final process, which stands for identifying and predicting model performance degradation, data skews and outliers. The central element of this lifecycle is data and model management, which are main functions for governing ML artifacts to support auditability, traceability, compliance, shareability, reusability, and discoverability of ML assets. An artifact is a discrete entity or piece of data produced and consumed by an ML workflow, such as datasets, models, input files and training logs. Note that data and model management processes are “transitional” phases. With the traditional components of an ML pipeline, you need to prepare data, perform feature engineering, train and tune the model, upload and store the model, compare the model to existing model versions, deploy the model, send prediction requests to the endpoint, present the model to any edge devices, and then monitor and manage the model. Vertex AI automates some of these steps, such as data preparation, feature engineering, model serving, and the ability to deploy to edge devices. Before you see how Vertex AI automates these steps, let’s first look at the reasons why Vertex AI serves the best in these different roles. Recall the different users within an organization that play a part in the ML lifecycle, such as product managers, data analysts and data engineers. These users have different needs and a comprehensive platform such as Vertex AI helps all of them thanks to its integration with different Google Cloud services that experiment with ML model development and quickly deploy solutions. There are four main reasons why Vertex AI serves the best in these different roles. First it provides a unified data and ML platform. A unified data and ML platform helps to build tighter connection points between data and ML. Therefore, it helps ML practitioners gain even greater value when they use all of our services to solve their largest problems. The second, Vertex AI lets ML practitioners achieve end-to-end MLOps. End-to-end MLOps helps ML practitioners efficiently and responsibly manage, monitor, govern, and explain ML projects throughout the entire development lifecycle. Thanks to its ready to use, prebuilt components, ML practitioners can achieve many ML tasks, including: Creating a new dataset and loading different data types, such as image, tabular, text, or video into the dataset. Exporting data from a dataset to Cloud Storage. Using BigQuery ML, Vertex AI custom training, or Vertex AI AutoML to train a model by using image, tabular, text, or video data. Running a custom training job by using a custom container or a Python package. Uploading an existing model to Vertex AI for batch or online prediction. And monitoring. These prebuilt components offer easier debugging and consistent interfaces to use standardized artifact types for input and output and cost efficiencies. Vertex AI also provides flexibility by delivering open and scalable ML infrastructure. The flexibility of data resources, ML framework, and hardware might accelerate the velocity of models into production. And finally, Vertex AI brings all the resources of Google to open source. In recent years, computing has both expanded as a field and grown in its importance to society. Similarly, the research conducted at Google has broadened dramatically. More than 3,000 Google and Deepmind researchers and more than 7,000 published publications drive better product performance and lower costs in many Google Cloud services recently. Building the next generation of data to AI, end-to-end MLOps and scalable infrastructure is happening in 3 phases: Phase 1 was for building a robust data foundation with services such as Bigtable or Pub/Sub. Phase 2 was for building an ML foundation with TensorFlow, TensorFlow Extended and Kubeflow. Phase 3 is now for building one unified AI platform by using many years of research and academic work. Vertex AI lets you access state-of-the-art AI algorithms developed by Google research to streamline complex AI use cases with optimized built-in infrastructure. Vertex AI drives more value from ML with access to state-of-the-art AI from Google Research and DeepMind and brings together all of Google Cloud services for building one unified AI platform with every ML tool you need, including the next generation of data-to-AI tools end-to-end MLOps capabilities, and scalable infrastructure. Vertex AI also integrates with widely used open source frameworks such as TensorFlow, PyTorch, and scikit-learn, along with supporting all ML frameworks and AI branches through custom containers for training and prediction.

2. Let's practice!

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.