As I was preparing my session for the recent DataOps Summit, I realized once again that data engineering is the future of data. More than that, those data engineers who rely on DataOps will lead the way. In this blog, as I did in my session, I will share why I believe that so strongly.
The Evolution to Modern Data Analytics
The purpose of data in the enterprise is to inform business decisions. And, of course, we do that with analytics. With the power of cloud computing, Business Intelligence (BI) that once only allowed reporting on past transactions has evolved into a diagnostic and descriptive analytics discipline that can operate on large amounts of data and at the speed of business. In addition, machine learning and AI enable predictive and prescriptive analytics to forecast sales accurately, understand and nurture their highest value customers, and more. In other words, it can help companies generate revenue and stay competitive.
It’s no wonder then that 451’s recent report, “Voice of the Enterprise: AI & Machine Learning Use Cases 2021,” found that 95% of enterprises surveyed consider AI to be important in their digital transformation efforts. It’s also not surprising to me that reality doesn’t live up to ideals.
IBM found just 21% of over 5000 companies surveyed had deployed AI, many AI PoCs never make it to production, and up to 70% of companies report no value from AI investments
But operationalization can turn these stats around.
Why Operationalization Matters
To demonstrate just why operationalization matters, let’s start with the granddaddy of all the Ops, DevOps. DevOps means a lot of things to a lot of people: agility, communication, collaboration, alignment, reliability, breaking down silos. These are great benefits of operationalization.
Without DevOps-style operationalization, ML models are often created in a silo. Disconnects can occur between the data science team creating the model and the IT team deploying it or the business team with the actual challenge. Your data science team may be working on perfecting something without enough business input due to an inability to test and deploy iterations continuously. And if it’s not right when it’s put into production? Well, if you’re not operationalized, you end up offline for weeks or months when you start all over again.
MLOps operationalizes this discipline and ensures that your model will evolve as change happens, without you having to stop, rework and start again.
And where reliability and agility are the benefits of operationalization, it’s the automation and monitoring that underpin operationalization that are the essence of XOps. Automation and monitoring overcome the huge dissonance between what people know and when and how they know it, providing the harmony that overcomes that cacophony between business, development, and operations.
Why MLOps and XOps Need DataOps
While operationalizing your models is a start, adding DataOps is the force multiplier for the effectiveness of your machine learning and MLOps. And it’s the same with any Ops discipline. You’ll find all of them, from CloudOps to SecOps to DevOps need DataOps.
To understand why we can look again at ML/AI. The more data an algorithm has to work with, the more accurate the results. But AI value, ML value, and analytics value are meaningful only if the data it operates on is valid across the whole ML lifecycle. You need sample data for exploration, test and training data for experimentation, and production data for evaluation. Traditional data integration methods may be capable of data quality procedures to ensure that only the cleanest data made it into the models — but those pipelines were brittle. The scale and complexity of today’s dynamic data architectures make this approach very risky. So as companies operationalize ML, they increasingly depend on smart data pipelines and DataOps, where data observability and pipeline resiliency are built into the pipelines themselves.
And it’s the same with all the Ops disciplines because they all need smart data pipelines. Beyond building, smart data pipelines must operate continuously. And so, we come to the three principles that fuel XOps success.
How DataOps Fuels XOps Success
Every Ops discipline needs continuous data and delivering that continuous data requires DataOps. The three key principles that allow the continuous delivery of data are continuous design, continuous operations, and continuous data observability.
Continuous design means that your data team can very easily start, extend, and collaborate on data pipelines on an ongoing basis. They do this with 10x less wasted time and 50x less downtime. It’s intent-driven so data engineers can focus on what they’re doing rather than how it’s being done. Continuous design is componentized so fragments of a pipeline can be reused as much as possible. And finally, there’s a single experience for every design pattern.
NatWest hosted a session highlighting advances in continuous design, incorporating repeatable patterns for ingestion to create a federated data engineering culture.
Continuous operations allow your data team to handle breakage easily, make a shift to new cloud platforms and respond to changes, whether breakages or business requests. It allows for automated deployments, with pipelines orchestrated in any mix and match combination of on-premises and cloud infrastructures and platforms. And most importantly, these data pipelines are decoupled as much as possible – within the pipeline, across pipelines, from origins, destinations, and external processes. The more you decouple, the easier it is to change.
DNB’s session highlighted how they’re perfecting the art of continuous operations with 20 engineers, enabling 200+ self-service data scientists and analysts for real-time for fraud detection.
Continuous data observability helps the data team understand the contents of the data and adhere to governance and compliance policies. It eliminates blind spots with a single, always-on Mission Control panel. Understanding data is priceless, essential to digital transformation and driving innovation.
BT shared how they practice continuous data observability at scale, monitoring over 10,000 pipelines with a single pane view of both on-premises and multiple clouds.
The Future of Data: The Means Not the Ends
The future of data is one where all its characteristics will be emergent. And what I mean by that is that you will obtain a macro understanding of your data by monitoring the emergent patterns in how people use the data, even as it evolves. The “ends” on the business value is not the result of a top-down process, where some anointed experts get together to understand the meaning of data and tell you how to implement data pipelines a priori. Instead, it is that of self-organizing patterns that result from cooperation between all of these autonomous micro participants. The key to this cooperation is an adherence to the same set of rules and principles. That is the “means” of DataOps.
And I’ll leave you with this: If you’re a data consumer, demand operationalization. If you’re a data provider or a data engineer, deliver operationalization. That’s the only way that data is going to become the lifeblood of business.