peshkova - Fotolia
As artificial intelligence becomes ubiquitous in the 21st century, IT ops pros have begun to wonder if and when...
they'll be automated out of a job.
To some extent, it's already begun to happen.
IT infrastructures are already automated and self-scaling, and self-healing systems are not far off as container orchestration tools develop. A cascade of IT monitoring tools with artificial intelligence and machine learning capabilities has hit the market promising to glue IT infrastructure with automated provisioning, deployment and incident response systems. IT pros can certainly be forgiven for asking if the future of DevOps has room for any humans.
The answer to that question, in these early stages of AIOps, depends on whom you ask. Most IT vendors are jumping wholeheartedly into an AI arms race, to the point where Microsoft and Amazon have declared that AI will be the crux of their future competitive advantage. The tech giants are so serious about the emerging field that they even collaborated on an AI-as-a-service product called Gluon. Google is also focused on AI, from its internal IT processes to its TensorFlow AI services.
The future of DevOps products will be AI-driven. Continuous integration and continuous delivery pipeline tools such as Electric Cloud 8.0 now incorporate machine learning and data analytics features to optimize DevOps workflows, and tools such as ServiceNow's Agent Intelligence bring machine learning to IT service ticket routing.
Nuno PereiraCTO, IJet International
Bleeding-edge DevOps shops already imagine a world of self-provisioning infrastructure to support application deployments.
"Why should I have to set up the resource requirements for each Kubernetes deployment?" said Cole Calistra, CTO of Kairos AR Inc., a provider of human facial recognition and analytics for developers in Miami. "Let the server figure out those limits based on actual historical data and predict what size the cluster should scale up to, based on what it has learned about its operation over time."
Many in the IT industry argue this is the only way IT can maintain staggeringly complex infrastructures of containers and virtual networks that support equally complex microservices architectures.
"By the end of 2018, we should see products closing the loop between data feeds from monitoring systems and orchestrators that take action," said Arvind Soni, VP of product at Netsil, a startup that just emerged from stealth with a tool that automatically maps and monitors Kubernetes infrastructures. "Container environments are so complex that anything else is unsustainable."
The doomsday proposition: AIOps replaces humans
IT ops pros are right to suspect the AI hype portends massive changes to their careers, regardless of whether AI replaces them entirely. And their worst-case scenario, that robots will completely take over their jobs, has some basis in reality.
Companies already claim products that can reduce the number of human operators needed to manage massive infrastructures. HCL Technologies Ltd, a multinational company based in India, said its ElasticOps product already applies AIOps to help maintain its managed cloud infrastructure service, a 50,000-instance environment, with just 30 engineers.
Machine learning algorithms and AI-assisted automation are the end goal for such tools as Datapipe Inc.'s Trebuchet application deployment platform, predicts Patrick McClory, SVP of platform engineering and delivery services at the managed cloud hosting services company based in Jersey City, N.J.*
"There's a strong undercurrent of protectionism around operations today in terms of, 'Don't automate my job,'" McClory said. But that protectionism won't shore up IT careers against AIOps for long.
"IT operations [is] a target of this, but applications are the thing that adds value to the business -- nobody really cares about infrastructure these days," he said. "Wouldn't it be cool if we go further up the stack instead of just instrumenting the behavior of these machines, to actually diving into the behavior of the developers working on it?"
If AIOps can realize that vision, humans will be back in the decision seat for a strategic role within companies, rather than caught up in undifferentiated day-to-day maintenance, McClory said.
Tale as old as time: AI technology and unintended consequences
As with any new technology, the early days of AI have already yielded unintended consequences that send shivers down the spine of a generation raised on movies such as Terminator and The Matrix, which present worst-case scenarios of machine intelligence run amok.
In the real world, Facebook's R&D staff was forced to pull the plug on an experiment this year that involved generative adversarial networks -- computer networks that can negotiate with one another -- when AI machines began speaking a language humans couldn't understand. This development was far from making Skynet a reality, but that didn't stop media outlets from pointing out that potential.
Early attempts to harness AI for IT management at Google also had real, if less dramatic, unintended consequences, according to Ben Sigelman, who served as senior staff software engineer for the web giant from 2003 to 2012.
Ben SigelmanCEO and co-founder, LightStep
"I saw things that correctly predicted almost every failure at Google and incorrectly predicted five times as many that weren't [accurate]," said Sigelman, who is now CEO and co-founder of LightStep, a startup that specializes in monitoring cloud-native microservices infrastructures. "It's incredibly powerful technology and can find signals in really noisy voluminous streams of information, but it needs to be the right signal. You shouldn't take any sort of action unless you're almost entirely certain that you're correct."
Discerning the right signals entirely depends on the data with which an AIOps system makes its decisions, and some experts argue that IT monitoring data collected to date isn't good enough to reinforce critical production environments.
"Data doesn't speak for itself -- we don't have enough or good enough data to generate models that we can really rely on," said Neil Raden, analyst at Wikibon. "Unattended machine learning algorithms just sift through data, and who knows if that's really effective."
Raden's colleague at Wikibon, James Kobielus, a former IBM AI evangelist that worked with the Watson AI platform, disagreed that AI doesn't have enough to go on, but acknowledged that human operators need to train AI algorithms on whether statistical correlations are valuable to the business.
But does reliance on human operators to train AI bring the field back to step one? The value of AI, after all, is to analyze much larger amounts of data than humans can handle, and potentially identify patterns humans can't.
For enterprises, early experiments in automatically generated IT monitoring alerts resulted in a wall of noise, which human operators quickly stemmed by paring down the number of alerts they received.
"The real question is whether we've overtuned it and now it's keeping some of the hidden gems hidden," said Nuno Pereira, CTO of IJet International, a risk management company in Annapolis, Md.
The company experienced a near-revolt among its ops team last year when IT monitoring data was hooked up to an automated paging system that flooded them with alerts. "That's one thing that keeps me up at night," Pereira said. "Are those needles in the ever-growing haystack being silenced?" As a result he's looking at AIOps tools from AppDynamics, which he already uses for other purposes, as well as competitors such as Moogsoft.
In the current furor around AIOps, as with any marketing buzzword, people latch onto a term but find it difficult to locate the signal in the noise, Pereira said. But as the volume of infrastructure data continues to increase, humans will inevitably need digital assistance, and he can't stop thinking about the needle in the haystack that may lie in wait.
The future of DevOps endgame: AIOps takes shape
How far can IT automation go? For now, even the most ardent AI supporters concede that only human-supervised AI is practical for use in IT shops in the next five to eight years.
"Our goal is not to just mature the technology and be confident in it from a statistical perspective, but also to work with the people to be more comfortable interacting with it, and to confirm or correct our assumptions around what should be done," Datapipe's McClory said.
Netsil's Soni foresees a "human augmentation" phase for AI that's already playing out in self-driving cars.
"We have the AI technology for a completely autonomous car, but would you trust it to drive your kids to school tomorrow?" he said. "Probably not. So what we have right now are augmentation features like blind spot and pedestrian warnings. The problem is not the technology, it's trust."
IT shops, which already have trouble finding skilled staff, must contend with a paradox as they look to AI to keep up with the future of DevOps. This technology could bridge the gap between overloaded IT staff and extremely complex modern infrastructures, but skills to develop and train machine learning are in short supply.
Credit bureau Experian, for example, is already deeply invested in AI and machine learning, especially in its R&D department, Experian DataLabs. The IT team at Experian also has AIOps on its radar, and the company has bots that automate some of its finance processes, said Barry Libenson, CIO at Experian. But while it's eager to expand on that, finding people to train AI systems is much easier said than done.
"We're constrained by the number of people we have with the expertise to do this stuff because it is so new," Libenson said. "Those skill sets are considerably more difficult to get, and can be considerably more complex than some of the stuff that's going on in the DevOps area."
*Information updated after publication
Automate infrastructure with AIOps monitoring tools
CA Automic prepares for increased DevOps spending with new tools
DevOps automation helps combat ongoing skills shortage problem
Is AI the key to root cause analysis?