I think it was physicist David Bohm who, in his book, ‘Wholeness and the Implicate Order‘ stated that “good theories, unlike bad ones, are true, but only up to a certain point“. In other words, it’s possible to be right about something, but not as completely right as you may think; to dramatically condense Hegel, reality is composed of many factors, all moving in relation to each other.
Which brings me to this 13 January Twitter post by the Duckbill Group’s Corey Quinn:
It’s not wrong to say that the “pile of VMs” estates Quinn describes form a significant part (what percentage, who can say?) of the usage profile for Azure customers. It’s not wrong, but it’s not complete. Indeed, I’d argue that the “pile of VMs” situation applies to all the cloud platforms when it comes to enterprise modernization inasmuch as corporate IT departments, under pressure to migrate workloads but typically not given time to refactor for cloud native architectures, often deploy familiar patterns to meet aggressive timelines (how many design decisions are made to ensure a CIO’s ‘we’re leaving the data center in 2 years‘ goal is realized?)
There are, however, specific factors in play for Azure customers – both organizational and practical – that shape usage. In my work as an architect, focused on Azure for years (almost since it has been a serious competitor), I’ve had the opportunity to gather empirical data – including the all revealing billing data – from a variety of businesses covering a broad range of categories – everything from materials science to fashion retail – and I think I can identify two, primary factors:
One.) Microsoft’s prominence, on premises, in the server-based space, whether bare metal or hypervisor.
It’s not uncommon for enterprise customers to host many thousands of Windows servers on premises. When the decision is made to migrate workloads hosted on these servers to Azure, IT departments tend to leverage the hybrid model or simply replatform existing architectures, something Microsoft has, if not encouraged exactly, facilitated via the way products such as Windows server have been engineered to be Azure-aware. Technology choices aside, there are almost always practical problems such as the challenge of performing an application modernization project for an on-premises estate spanning across 15,000 VMs (yes, I’m describing a real project) in very short time periods. These factors, of necessity, funnel efforts into creating “pile of VMs” style solutions.
Two.) Azure Workloads Are Infrastructural (but not in the way you think)
From time to time (usually when there’s been a cloud outage such as the 7 January AWS event) Quinn asserts that Azure probably isn’t used for high profile workloads such as customer-facing content delivery – Netflix, HBOMAX, et. al. – because, if Azure was a large part of that market, outages on the platform would result in downtime for these prominent global services (a classic syllogism not unlike: all men are mortal, Socrates is a man therefore, Socrates is mortal).
This isn’t wrong but the conclusion Corey typically reaches – that Azure is only, or is principally used for back office workloads such as SharePoint, etc. is an example of, as Bohm might have said, going beyond the effective bounds of your theory and wandering into error.
In my experience, Azure workloads are increasingly providing large-scale service platforms in three areas:
- IoT – for example, I designed a solution for a manufacturer who uses Azure IoT to gather data from partner/customers to determine in near real time the state of products – in this case, compressed gases distributed across thousands of venues in North America, Europe and Asia)
- Data Analytics – more and more of my customers have deployed or are deploying analytics services such as Databricks and Synapse Analytics. These offerings are so popular they’re forming a large percentage of bills – I’ve written about this here. For example, I worked with one of my customers to build a Databricks environment that uses machine learning to vectorize the chemical formulas which form the basis of their product (these formulas are statistically assessed alongside sales and location data to determine where research and sales efforts might be best directed).
- AI and Machine Learning – there’s lots of hype about these areas and a lot of marketing nonsense. At a practical level, I’m seeing customers putting aspects of the Azure Cognitive Services API suite to real-world use, building applications that leverage the following models:
- Vision processing
- Natural Language Processing
- Decision support
For example, I recently worked with a customer to architect and deploy a vision processing solution which uses Azure Forms Recognizer to increase the velocity of processing standard forms.
Beyond my direct experience, there are projects such as this Siemens analytics and anomaly detection initiative (I know some of the people involved so it’s not just marketing hype) this Bosch project and many others that demonstrate the foundational, business systems critical nature of deployed Azure solutions.
These solutions are infrastructural in the sense that they are designed to provide the structure that supports more visible business process.
Expanding the Map to Visualize More of the Territory
It is indeed true that “piles of VMs” are a significant part of the story of how Azure is used. To the extent this is true, it’s the result of organizational habits and practical challenges that are difficult for many enterprises to overcome – at least rapidly. Beyond this ‘pile’ however, there’s a broad range of activities underway that, because of their foundational nature, are hidden from everyone who isn’t directly working with clients to build Azure solutions. Companies don’t often talk about these projects and customers rarely see them. Even so, they are a significant part of Azure’s usage profile.
Azure FinOps 