MLOps - Machine Learning Operations
What is MLOps?
MLOps (machine learning operations) is a practice that aims to make developing and maintaining production machine learning seamless and efficient. While MLOps is relatively nascent, the data science community generally agrees that it’s an umbrella term for best practices and guiding principles around machine learning – not a single technical solution.
The Guiding Principles of MLOps
Machine learning should be collaborative.
When approached from the perspective of a model – or even ML code, machine learning cannot be developed truly collaboratively as most of what makes the model is hidden. MLOps encourages teams to make everything that goes into producing a machine learning model visible – from data extraction to model deployment and monitoring. Turning tacit knowledge into code makes machine learning collaborative.
Machine learning should be reproducible.
Data scientists should be able to audit and reproduce every production model. In software development, version control for code is standard, but machine learning requires more than that. Most importantly, it means versioning data as well as parameters and metadata. Storing all model training related artifacts ensures that models can always be reproduced.
Machine learning should be continuous.
A machine learning model is temporary. The lifecycle of a trained model depends entirely on the use-case and how dynamic the underlying data is. Building a fully automatic, self-healing system may have diminishing returns based on your use-case, but machine learning should be thought of as a continuous process and as such, retraining a model should be as close to effortless as possible.
Machine learning should be tested & monitored.
Testing and monitoring are part of engineering practices, and machine learning should be no different. In the machine learning context, the meaning of performance is not only focused on technical performance (such as latency) but, more importantly, predictive performance. MLOps best practices encourage making expected behavior visible and to set standards that models should adhere to, rather than rely on a gut feeling.
Do you think something is missing from the above, email us with your opinion.
As mentioned before, MLOps is not dependent on a single technology or platform. However, technologies play a significant role in practical implementations of MLOps, similarly to how adopting Scrum often culminates in setting up and onboarding the whole team to e.g. JIRA. Therefore, the project to rethink machine learning from an operational perspective is often about adopting the guiding principles and making decisions on infrastructure that will support the organization going forward.
Practical MLOps: How to Get Ready for Production Models
This eBook is a definitive guide to the concepts of machine learning operations.
The MLOps Stack
This template helps visualizes your machine learning infrastructure stack and consider tooling choices.
Why MLOps Matters?
The goal of MLOps is to reduce technical friction to get the model from an idea into production in the shortest possible time to market with as little risk as possible.
Traditionally, machine learning has been approached from a perspective of individual scientific experiments predominantly carried out in isolation by data scientists. However, as machine learning models become part of real-world solutions and critical to business, we will have to shift our perspective, not to depreciate scientific principles but to make them more easily accessible, reproducible, collaborative, and most importantly to increase the speed at which machine learning capabilities can be released.
The reality is that only a model running in production can bring value. Models have zero ROI until they can be used. Therefore, time to market should be the number one metric to look at and optimize for any commercial ML project.
Priorities in ML for the next 3 monthsSource: State of ML 2020, 330 respondents
In May 2020, we surveyed 330 data scientists, machine learning engineers, and managers in a broad range of companies to ask what they were focused on for the next three months and what major obstacles they were faced with. Half of the respondents said that they were focused on developing models for production use, and over 40% said they would be deploying models to production.
Automating retraining of models and monitoring models in production were still not relevant for most respondents, which speaks to the practice of MLOps of being still relatively nascent. Production models raise new challenges, not just for data scientists but for the extended team of engineers, product managers, and compliance officers, which will need to be solved collaboratively.
In most real-world applications, the underlying data changes constantly, and thus models need to be retrained, or even whole pipelines need to be rebuilt to tackle feature drift. Business and regulatory requirements can also change rapidly, requiring a more frequent release cycle. This is where MLOps comes in to combine operational know-how with machine learning and data science knowledge.
How to Get Started with MLOps?
Step 1: Recognize the stakeholders.
The size and scope of real-world machine learning projects have surely surprised most – if not all of us. What seems like a straightforward task of gathering some data, training a model and then using it for profit ends up becoming a deep rabbit hole that spans from business and operations to IT. A single project covers data storage, data security, access control, resource management, high availability, integrations to existing business applications, testing, retraining, etc. Many machine learning projects end up being some of the biggest multidisciplinary and cross-organizational development efforts that the companies have ever faced.
To properly implement an MLOps process, you’ll have to recognize the key people and roles in your organization. We’ve talked to countless organizations in the past four years, and while each case is unique, there tends to be a combination of these roles that contribute machine learning projects:
- Data scientists (duh!)
- Data engineers
- Machine learning engineers
- DevOps engineers
- Business owners
These roles are not necessarily one per person, but rather a single person can - and in smaller organizations often has to - cover multiple roles. However, it paints a picture of who you’ll need to identify to gather your specific requirements and use your organization’s resources. For example, you must talk to business owners to understand the regulatory requirements and IT to grant access and provision cloud machines.
Step 2: Invest in infrastructure.
MLOps is all about operational infrastructure that makes developing machine learning more efficient. There are plenty of proprietary and open-source products that solve parts of the machine learning lifecycle. When comparing platforms in the MLOps space, you’ll often run into apples to oranges comparisons. For example, comparing KubeFlow and Valohai is tricky because the former is an extendable, open-source solution requiring weeks to adopt, and the latter is a managed, proprietary solution.
To make it more straightforward on how to decide on what infrastructure solutions to adopt, you should consider the following aspects:
- Reproducibility Will the solution make it easier to retain knowledge about data science work?
For example, experiment tracking and data version control are solutions that will significantly increase your teams capability to reproduce previous work.
- Efficiency Will the solution save us time or money?
For example, automated machine orchestration will reduce the risk of paying for costly GPU instances even when they are not in use, and pipeline capabilities will remove manual work from routine operations.
- Integrability Will the solution integrate with our existing systems and processes?
For example, feature store will make it easier for data engineers and data scientists to work together around data and integrate with ML platforms.
There are many approaches to ML infrastructure that can work, whether it’s coupling specialized systems or using a single multipurpose platform. However, infrastructure work should start as early as possible to avoid situations where your team has models in production, but how the models were produced isn’t well documented and each new release is becoming increasingly difficult to do.
Step 3: Automate, automate, automate.
When moving from POC to production, there is a significant change in mindset you’ll have to make. The ML model is no longer the product; the pipeline is.
While machine learning projects often start with one huge notebook and some local data, it’s essential to begin splitting the problem solving into more manageable components; components that can be tied together but tested and developed separately.
This ability to split the problem solving into reproducible, predefined and executable components forces the team to adhere to a joined process. A joined process, in turn, creates a well-defined language between the data scientists and the engineers and also eventually leads to an automated setup that is ML equivalent of continuous integration (CI) – a product capable of auto-updating itself.
An end-to-end, automated machine learning pipeline ensures that every change – in either data or code – is (or can be) deployed to production without it turning into a special project.
The result of MLOps should be a supercharged release cycle of machine learning capabilities. It’s paramount to understand that MLOps is a combination of people and technologies; people have to be willing to automate and document their work and tools have to make it easy.
For more MLOps advice, sign up for our eBook and receive your copy today.
Learn what MLOps is all about and how MLOps helps you avoid the deadlock between machine learning and operations. This eBook gives an overview of why MLOps matters and how you should think about implementing it as a standard practice.