Azure Machine Learning vs MLFlow

Machine learning and deep learning models are difficult to get into production due to the highly cyclical development process of data science. As a data scientist I am not only experimenting with different hyper-parameters to tune my model, but I am constantly creating or adding new features in my dataset; therefore, needing to keep track of which features, hyper-parameters, and the type of model used can be a lot to remember. Additionally, I have to create a model that performs up to our evaluation criteria, otherwise, the process of deploying a model is irrelevant. While their seems to be an endless number of tools that can help engineers track and monitor tasks throughout the development process, tools like MLFlow and Azure Machine Learning look to help make this process manageable.

Once a data scientist develops a machine learning model, the code is rarely production ready and package dependencies are usually a mess. Not only do developers need to clean up the model training script, but they typically need to make big changes in how they clean and acquire data. During the training process, data scientists are able to access cold historical data that is easy to acquire and test on. However, once in production we need new code that acquires data as it is created, transforms it as needed, and make predictions. In addition, to a new data acquisition script data scientists need to write a new scoring script that actually makes predictions and returns those predictions to the application. In the end, data science solutions have a lot of moving parts in development and in production.

For the longest time I would recommend Azure Machine Learning to my clients to track, train, and deploy machine learning solutions. However, as Azure Databricks has become the premier big data and analytics resources in Azure, I have had to adjust that message slightly. Since MLFlow is integrated into Azure Databricks it has easily become the default platform to manage data science experiments from development to production in a Spark environment, however, I believe that Azure Machine Learning is a viable, and often better, tool choice for data scientists. In the demonstration available on my GitHub I show users how to train and track machine learning models using MLFlow, Azure Machine Learning, and MLFlow’s integration with Azure Machine Learning. If you are looking to deploy models using Azure Databricks and Azure Machine Learning, check out my previous demo available in the same repository.

MLFlow is centered around enhancing the engineer’s ability to track experiments so that they have visibility of performance during both development and production. Managing models is simplified by associating them with specific experiment runs, and it packages machine learning code so that it is reusable, reproduceable, and shareable across an organization. With MLFlow a data scientist is able to execute and compare hundreds of training runs in parallel on any compute platform they wish and deploy those models in any manner they desire. MLFlow supports all programming languages through its REST interface, however, R, Python, Java, and Scala are all supported languages out of the box. Below is a screenshot of a MLFlow experiment view in Azure Databricks.

If I were to select one of my experiment runs by clicking on the execution date hyperlink then I can see more details about the specific run. We also have the ability to select multiple runs and compare them.

Azure Machine Learning is an enterprise ready tool that integrates seamlessly with your Azure Active Directory and other Azure Services. Similar to MLFlow, it allows developers to train models, deploy them with built in Docker capabilities (please note that you do not have to deploy with Docker), and manage machine learning models in the cloud. Azure machine learning is fully compatible with popular packages like PyTorch, TensorFlow, and scikit-learn, and allows developers to train models locally and scale out to the cloud when needed.

At a high-level the Azure Machine Learning Service provides a workspace that is the central container for all development and artifact storage. Within a workspace a developer can create experiments that all scripts, artifacts, and logging is tracked by the usage of experiment runs. The most important aspect of data science is our model, the object that takes in our source data and makes predictions. Azure Machine Learning provides a model registry that tracks and versions our experiment models making it easier to deploy and audit predictive solutions. One of the most crucial aspects to any machine learning solution is deployment. The Azure Machine Learning service allows developers to package their python code as a web service Docker container. These docker images and containers are cataloged an Azure Container Registry that is associated to the Azure Machine Learning Workspace. This give data scientists the ability to track a single training run from development into production by capturing all the training criteria, registering our model, building a container, and creating a deployment.

Below is a list of my experiments in a demo Azure Machine Learning Workspace.

By clicking into one of the experiments I am able to see all the runs and view the performance of each run through the values logged using the AzureML Python SDK. Please note that I have the ability to select multiple runs and compare them.

The above screenshots are very similar to MLFlow, where I believe Azure Machine Learning extends and offers better capabilities is through the Compute, Models, Images, and Deployment tabs in our Azure ML Workspace.

Either programmatically or using the Azure Portal I am able to create a remote compute target where I can offload my experiment runs from my local laptop, and have everything logged and stored in my workspace.

By registering models in my workspace I make them available to create a Docker image and deploy as a web service. Developers can either use the Azure ML SDK or the Azure portal to do so.

Once an image is created I can easily deploy the Docker container anywhere that I can run containers. This can be in Azure, locally, or on the edge! One extremely nice feature built into Azure Machine Learning is the integration with Application Insights allow developers to capture telemetry data about the web service and the model in production.

Overall, while MLFlow and Azure Machine Learning are very similar, I typically side with Azure Machine Learning as the more enterprise ready product that enables developers to deploy solutions faster. However, the cross validation ability that is built into MLFlow, mllib, and Databricks makes it extremely easy to tune hyper-parameters, while the Azure Machine Learning hyper-parameter tuning is a little more difficult.

One my favorite features of MLFlow and Azure Machine Learning is the ability to use MLFlow in union with Azure Machine Learning, which I highlight in my demo. Generally, I recommend to engineers who are developing exclusively on Azure Databricks to use MLFlow due to the easy integration it provides, however, if there is a subset of solutions being deployed or developed in a non-Spark environment I would recommend a tool like Azure Machine Learning to centralize all data science experiments in one location. Please check out the example of MLFlow and Azure Machine Learning on Azure Databricks available on my github!

Quick Review: Databricks Delta

As the number of data sources grow and the size of that data increases, organizations have moved to building out data lakes in the cloud in order to provide scalable data engineering workflows and predictive analytics to support business solutions. I have worked with several companies to build out these structured data lakes and the solutions that sit on top of them. While data lakes provide a level of scalability, ease of access, and ability to quickly iterate over solutions, they have always fallen a little short on the structure and reliability that traditional data warehouses have provided.

Historically I have recommended that customers apply structure, not rules, to their data lake so that it makes the aggregation and transformation of data easier for engineers to serve to customers. The recommended structure was usually similar lambda architecture, as not all organizations have streaming data, but they would build out their data lake knowing this was a possibility in the future. The flow of data generally followed the process described below:

• Batch and streaming data sources are aggregated into raw data tables with little to no transforms applied i.e. streaming log data from a web application or batch loading application database deltas.
• Batch and streaming jobs in our raw data tables are cleaned, transformed, and saved to staging tables by executing the minimum number of transforms on a single data source i.e. we tabularize a json file and save it as a parquet file without joining any other data or we aggregate granular data.
• Finally we aggregate data, join sources, and apply business logic to create our summary tables i.e. the tables data analysts, data scientists, and engineers ingest for their solutions.

One key to the summary tables is that they are business driven. Meaning that we create these data tables to solve specific problems and to be queried on a regular basis. Additionally, I recently took a Databricks course and instead of the terms raw, staging, and summary; they used bronze, silver, and gold tables respectfully. I now prefer the Databricks terminology over my own.

Delta Lake is an open source project designed to make big data solutions easier and has been mostly developed by Databricks. Data lakes have always worked well, however, since Delta Lake came onto the scene, organizations are able to take advantage of additional features when updating or creating their data lakes.

• ACID Transactions: Serial transactions to ensure data integrity.
• Data Versioning: Delta Lake provides data snapshots allowing developers to access and revert earlier versions of data for audits, rollbacks, and reproducing predictive experiments.
• Open Format: Data stored as in Parquet format making it easy to convert existing data lakes into Delta Lakes.
• Unified Batch and Streaming: Combine streaming and batch data sources into a single location, and use Delta tables can act as a streaming source as well.
• Schema Enforcement: Provide and enforce a schema as need to ensure correct data types and columns.
• Schema Evolution: Easily change the schema of your data as it evolves over time.

Generally, Delta Lake offers a very similar development and consumption pattern as a typical data lake, however, the items listed above are added features that bring an enterprise level of capabilities that make the lives of data engineers, analysts, and scientists easier.

As an Azure consultant, Databricks Delta is the big data solution I recommend to my clients. To get started developing a data lake solution with Azure Databricks and Databricks Delta check out the demo provided on my GitHub. We take advantage of traditional cloud storage by using an Azure Data Lake Gen2 to serve as the storage layer on our Delta Lake.

Azure Machine Learning Services and Azure Databricks

As a consultant working almost exclusively in Microsoft Azure, developing and deploying artificial intelligent (AI) solutions to suit our client’s needs is at the core of our business. Predictive solutions need to be easy to implement and must scale as it becomes business critical. Most organizations have existing applications and processes that they wish to infuse with AI. When deploying intelligence to integrate with existing applications it needs to be a microservice type feature that is easy to consume by the application. After trial and error I have grown to love implementing new features using both the Azure Machine Learning Service (AML Service) and Azure Databricks.

Azure Machine Learning Service is a platform that allows data scientists and data engineers to train, deploy, automate, and manage machine learning models at scale and in the cloud. Developers can build intelligent algorithms into applications and workflows using Python-based libraries. The AML Service is a framework that allows developers to train wherever they choose, then wrap their model as a web service in a docker container and deploy to any container orchestrator they wish!

Azure Databricks is a an optimized Apache Spark Platform for heavy analytics workloads. It was designed with the founders of Apache Spark, allowing for a natural integration with Azure services. Databricks makes the setup of Spark as easy as a few clicks allowing organizations to streamline development and provides an interactive workspace for collaboration between data scientists, data engineers, and business analysts. Developers can enable their business with familiar tools and a distributed processing platform to unlock their data’s secrets.

While Azure Databricks is a great platform to deploy AI Solutions (batch and streaming), I will often use it as the compute for training machine learning models before deploying with the AML Service (web service).

Ways to Implement AI

The most common ways to deploy a machine learning solution are as a:

• Consumable web service
• Scheduled batch process
• Continuously streaming predictions

Many organizations will start with smaller batch processes to support reporting needs, then as the need for application integration and near real-time predictions grow the solution turns into streaming or a web service.

Web Service Implementation

A web service is simply code that can be invoked remotely to execute a specific task. In machine learning solutions, web services are a great way to deploy a predictive model that needs to be consumed by one or more applications. Web services allow for simply integration into new and existing applications.

A major advantage to deploying web services over both batch and streaming solutions is the ability to add near real-time intelligence without changing infrastructure or architecture. Web services allow developers to simply add a feature to their code without having to do a massive overhaul of the current processes because they simply need to add a new API call to bring those predictions to consumption.

One disadvantage is that predictions can only be made by calling the web service. Therefore, if a developer wishes to have predictions made on a scheduled basis or continuously, there needs to be an outside process to call that web service. However, if an individual is simply trying to make scheduled batch calls, I would recommend using Azure Databricks.

Batch Processing

Batch processing is a technique to transform a dataset at one time, as opposed to individual data points. Typically this is a large amount of data that has been aggregated over a period of time. The main goal of batch processing is to efficiently work on a bigger window of data that consists of files or records. These processes are usually ran in “off” hours so that it does not impact business critical systems.

Batch processing is extremely effective at unlocking *deep insights* in your data. It allows users to process a large window of data to analyze trends over time and really allow engineers to manipulate and transform data to solve business problems.

As common as batch processing is, there are a few disadvantages to implementing a batch process. Maintaining and debugging a batch process can sometimes be difficult. For anyone who has tried to debug a complex stored procedure in a Microsoft SQL Server will understand this difficulty. Another issue that can arise in today’s cloud first world is the cost of implementing a solution. Batch solutions are great at saving money because the infrastructure required can spin up and shut down automatically since it only needs to be on when the process is running. However, the implementation and knowledge transfer of the solution can often be the first hurdle faced.

By thoughtfully designing and documenting these batch processes, organizations should be able to avoid any issues with these types of solutions.

Stream Processing

Stream processing is the ability to analyze data as it flows from the data source (application, devices, etc.) to a storage location (relational databases, data lakes, etc.). Due to the continuous nature of these systems, large amounts of data is not required to be stored at one time and are focused on finding insights in small windows of time. Stream processing is ideal when you wish to track or detect events that are close in time and occur frequently.

The hardest part of implementing a streaming data solution is the ability to keep up with the input data rate. Meaning that the solution must be able to process data as fast or faster than the rate at which the data sources generate data. If the solution is unable to achieve this then it will lead to a never ending backlog of data and may run into storage or memory issues. Having a plan to access data after the stream is operated on and reduce the number of copies to optimize storage can be difficult.

While there are difficulties with a streaming data architecture, it enables engineers to unlock insights as they occur. Meaning, organizations can detect or predict if there is a problem faster than any other method of data processing. Streaming solutions truly enable predictive agility within an organization.

Check out the Walkthrough

Implementing a machine learning solution with Azure Databricks and Azure Machine Learning allows data scientists to easily deploy the same model in several different environments. Azure Databricks is capable of making streaming predictions as data enters the system, as well as large batch processes. While these two ways are great for unlocking insights from your data, often the best way to incorporate intelligence into an application is by calling a web service. Azure Machine Learning service allows a data scientist to wrap up their model and easily deploy it to Azure Container Instance. From my experience this is the best and easiest way to integrate intelligence into existing applications and processes!

Check out the walkthrough I created that shows engineers how to train a model on the Databricks platform and deploys that model to AML Service.

Streaming Machine Learning with Azure Databricks

Organizations are beginning to not only benefit from streaming data solutions, but require them to differentiate themselves from their competitors. Real-time reporting, alerts, and predictions are now common requests for businesses of all sizes.

That said, they rarely understand the requirements or implementation details needed to achieve that level of data processing. Streaming data is information that is generated and consumed continuously. This information typically includes many data sources, including log files, point of sale data (in store and online), financial data, and IoT Devices, to name just a few.

Implementation

Fast and Easy

Generating business-changing insights from streaming data can be a difficult process; however, there are quick wins for organizations of all sizes. Microsoft Azure offers an abundance of PaaS or SaaS products that allow users to connect to sources and automate workflows.

With Azure Logic Apps, it is extremely easy to set up data pipelines that extract data from your social media pages, analyze them for sentiment analysis, and alert users when comments or posts need to be addressed. While this may not be a business-changing solution, it gives companies the ability to have a more intimate level of interaction with customers or users than they had before.

Microsoft has provided a simple solution for companies to take advantage of this capability. Using Azure Logic Apps and Microsoft Cognitive Services, one can be alerted of any positive or negative tweets that occurs about their company. This is an easy and cost-effective way to implement intelligence into workflows. (Check out the example available here.) Azure Logic Apps connect to a variety of data sources, enabling organizations to obtain a quick win for real-time reporting with a deceptively simple drag-and-drop interface.

Ideal Implementation

From my experience, companies benefit most from custom machine learning solutions that solve a specific business problem using their own data. Creating solutions tailored to solve a problem in a specific environment allows a business to truly take a proactive approach as they incorporate intelligence throughout their organization. However, lack of knowledge is often a barrier for companies when implementing custom and scalable solutions.

Azure Databricks is an optimized Apache Spark platform perfect for data engineering and artificial intelligence solutions. It is an ideal platform for implementing batch or streaming processes on business critical data, and enables developers to create and deploy predictive analytics (machine learning and deep learning) solutions in an easy to use notebook environment.

Initially, organizations may implement their solutions as batch processes on Azure Databricks to save on cloud consumption costs, but plan for the future by using a platform that will scale and grow with the needs of the business. Batch processes allow users to save on monthly costs by turning off your virtual machines when they are not used, then when real-time insights is required the developer can almost flip a switch for streaming data. Deploy cost effective infrastructure now with the ability to scale limitlessly as you need in the future.

Below is a common infrastructure diagram I implement with my customers. If streaming is not required then we simply bypass the event hub and write python or scala scripts to connect directly to the data sources.

1. A number of data sources (devices, applications, databases etc.) that publish information to an Azure Event Hub (or Apache Kafka).
   1. Please note that whatever the data source is, there will always need to be some sort of process or application that collects data and sends it to the Event Hub.
2. Azure Databricks will write the stream of data as quickly as possible to an unaltered, “raw”, data storage in an Azure Data Lake Store or Azure Blob Storage.
3. In addition to writing to raw storage, Databricks will be used to cleanse data as needed and stream appropriately to an application database, Power BI, or use Databricks Delta for real-time insights, consumption, and intelligent automated actions. Please note that applications can read directly off an Event Hub as a consumer as well.
4. Then use Azure Databricks to train a machine learning or deep learning model that can be used to make streaming or batch predictions.

Tips to Actually Implement a Solution

When implementing new intelligent solutions with cloud infrastructure, it is likely that it will require internal business stakeholder buy in. Therefore, in order to successfully implement a new predictive analytics solution you must:

1.  Identify a business problem to solve and the stakeholders
2.  Visualize or surface results to “wow” stakeholders
3.  Start developing iteratively

If a team attempts to solve too many problems initially by trying to answer all possible questions, they will likely fail to “wow” a business user. Developers will likely focus all their time on coding and analyzing the best path forward that they will only have code to show (code is a rather boring deliverable for most business users), and may simply never get past the proof of concept or analysis phase.

Business Problem

It is common for companies to simply start creating a solution to work with newer technology without a true business problem they are trying to solve. It happens most often for organizations who want to start a data lake strategy. Their main goal is to develop a data lake so that other business units can take advantage of the sandbox environment for predictive analytics.

I believe a centralized data lake for organizations is a great idea for any IT group. However, without a specific business problem, it is difficult to see the true value that a data lake or machine learning solution provides, which in turn can slow adoption. By focusing on solving a single use case other, there will be a reference to other business units on why they should use the enterprise data lake. The reason for adoption is much more tangible.

Wow Stakeholders

There is not a more boring outcome to a business stakeholder than a project resulting in code. Machine learning or deep learning projects must have some type of end product that accurately describes the effectiveness of the solution created. In most machine learning solutions that I implement, I will almost always provide a Power BI Report. This ensures that the model and predictions are tangible because they are shown through visualizations. The business user now has the ability to actually use the predictions and show other internal users the solution.

Iterative Development

The most frustrating part of projects can be the initial planning or analysis phase. Large enterprises will often start a project and get stuck in analysis paralysis. I encourage teams I work with to simply start coding! This does not mean to do zero planning or proof of concepts, but at some point a team has to pick a direction and run with it. Avoid over analyzing various products by picking a small subset of well-known products, analyze them, and go!

Benefits

Streaming data architecture is beneficial in most scenarios where dynamic data is generated on a continual basis. Any industry can benefit from data that’s available almost instantly from the time it was created. Most organizations will begin with simple solutions to collect log data, detect outliers based on set (unintelligent) rules, or provide real-time reporting.

However, these solutions evolve, becoming more sophisticated data processing pipelines that can learn and detect outlier data points as they occur. The true advantage of streaming data is in performing advanced tasks, like machine learning, to take preventive or proactive action.

Processing a data stream effectively generates quick insights, but it does not replace batch processes. Typically, organizations implement both solutions to gain quick, more computationally intensive insights. Streaming data reacts to or anticipates events, while batch processing derives additional insights after the fact.

Batch processing can often require more compute. It’s ideal when time or speed is not a priority. One of the biggest advantages of Azure Databricks is that companies are able to use the same infrastructure for both their workflows!

Batch processing data requires a system to allow data to build up so that it can be processed all at once. This often requires larger compute resources than streaming due to the size of data, which can be a hurdle for most organizations; however, it allows users to aggregate and analyze large amounts of data over a longer period of time. Streaming solutions do less computing, but require machines to be running 100% of the time and typically look at data over a shorter period of time.

Example

I recently created a simple walkthrough of how to implement a streaming data solution on Azure. Check out the walkthrough on GitHub. Please note that an Azure subscription is required.

Conclusion

Organizations of any size can benefit from a streaming solution using Databricks and Azure Data Lake Store. It enables near real-time reporting, as well as, provides a sandbox environment for iterative development of intelligent solutions. Azure Databricks and Data Lake Store allow a developer to implement both batch and streaming solutions in a familiar and easy to use environment.