Data Science 3.0 with MLOps

Implementing MLOps

MLOps is a relatively new term that has been cropping up more and more lately. And for good reason! Given the current state of data science, it can help make data science processes more effective, efficient, secure and reliable. That sounds promising, but what exactly is MLOps, how does it work and how can you get started? Or why should you care in the first place? We’ll explore these questions in this blog!

The state of data science

Data science is a fairly young field and is still developing in most organisations. Some organisations are further along than others, but there is a general pattern:

Phase 1: the lab

It starts with a lab: a workshop for alchemists who turn data into gold with complex algorithms (or so we think).

When it turns out that, apart from a few PowerPoints with interesting charts, the lab has yet to produce anything of value, the pressure starts to build: results are needed, something has to be delivered!

Phase 2: in production

At last, a solid business case has been found and, a few months later, there is actually a predictive model that works well. It then takes another 9 months of rewriting code and discussions with IT and architects (target landscape, supported tooling, OTAP, security) before it finally goes into production. But still: it’s there, it’s running and it works! Hooray!

Until, suddenly, it doesn’t. The predictions are off, users want something else, or it turns out the process broke down a month ago. But who is going to fix it? Jantje, who built the model, left long ago. Pietje cannot make sense of his undocumented notebooks and, after several attempts to fix things, concludes he might as well rebuild it ‘from scratch’ .

Phase 3: maturity

Having learnt the hard way, everything now runs like a well-oiled machine. Whether it’s retrieving, loading and sharing data, managing different model experiments, automatically deploying AI-driven APIs and apps with CI/CD, or explaining predictions… Every team member does it at the touch of a button and with a smile. Exactly: this is science fiction.

MLOps to the rescue

To summarise the scenario above a little more formally:

• Phase 1: Experimentation
  • Data science operates in isolation from Business and IT
  • Few concrete results
  • No real value delivered
• Phase 2: Go-live
  • Outputs are put into production
  • But management, monitoring, reproducibility, security, etc. are not yet in place
  • Initial value is delivered, but it is not sustainable
• Phase 3: Maturity
  • Processes are effective, efficient, reproducible and secure
  • Value is delivered continuously

That is where MLOps comes in: it is a way to reach phase 3.

What is MLOps?

“With Machine Learning Model Operationalization Management (MLOps), we want to provide an end-to-end machine learning development process to design, build and manage reproducible, testable, and evolvable ML-powered software.”

⎼ CDF sig-MLOps

“the extension of the DevOps methodology to include Machine Learning and Data Science assets as first class citizens within the DevOps ecology”

⎼ also CDF sig-MLOps

“MLOps empowers data scientists and app developers to help bring ML models to production. MLOps enables you to track / version / audit / certify / re-use every asset in your ML lifecycle and provides orchestration services to streamline managing this lifecycle.”

⎼ Microsoft

In short, you could say that MLOps is DevOps for machine learning / data science.

• In DS/ML/AI, we ultimately develop software too
• So we can apply DevOps
• But DS/ML/AI involves more than software development: think datasets, models and training results
• So DevOps needs an extension: MLOps

The aim of MLOps is:

Our data science processes are

• Effective
• Demonstrably effective
• Efficient (fast)
• Reproducible
• Secure
• Robust/reliable
• Easy to use

OK, sounds good, another magic formula… So if you shout MLOps, do all your data science problems magically disappear, just as agile was supposed to solve all your business problems?

What MLOps isn’t

Everything is bad  -> We introduce MLOps -> Everything is good.

So it’s not a magic spell.

• There are best practices, but…
• You need to set up MLOps for your own processes
• in line with your own needs and priorities

But let’s get specific

Let’s move beyond the high-level view. What processes are we talking about?

Data science life cycle