As more IT organisations apply artificial intelligence (AI), machine learning (ML), and so-called AIOps technology to network management, network data is critical to success. AI/ML technology requires more and more data to learn individual networks, derive insights, and offer recommendations. Unfortunately, many organisations encounter problems when trying to feed network data to these AI tools.
In other words, network teams need to modernise their approach to network data before they embrace AI technology.
Enterprise Management Associates recently surveyed 250 IT professionals about their experience with AI/ML-driven network management solutions for a report, “AI-Driven Networks: Leveling up Network Management.” It found that data problems are the number-two technical challenge they encounter when applying AI/ML to network management. Only network complexity is a bigger technical issue.
It also found that found that 90% of organisations have encountered at least one serious challenge with network data when trying to use their AI/ML solutions.
“AIOps needs data to drive its workflows,” an IT vice president with a $9 billion financial services company said recently. “If you don’t have data, you don’t have AIOps. The first thing you need to do [with an AI project] is get your data ready. Look at it, understand it, and see where the gaps are.”
Here are the key sources of data trouble, according to those IT pros surveyed.
The number one issue, affecting 46% of organisations, was data quality. IT organisations quickly discover that garbage data produces garbage insights. They’re struggling with errors, formatting issues, and nonstandard data. This can especially be an issue if an IT organisation is feeding data from multiple siloed tools into a third-party AIOps solution. The typical IT organisation uses anywhere from four to 15 tools to manage and monitor its network. Each tool maintains its own database with varying levels of quality. When an AIOps solution tries to correlate insights across those data sets, problems will emerge.
Nearly 39% told EMA that they are struggling with the security risk associated with sharing network data with AI/ML systems. Many vendors offer AI-driven networking solutions as cloud-based offerings. IT teams must send their network data into the cloud for analysis. Some industries, like financial services, are averse to sending network data into the cloud. They’d rather keep it in-house with an on-premises tool. Unfortunately, many network vendors won’t support an on-premises version of their AI data lake because they need cloud scalability to make it work.
Some vendors also combine the anonymised data of all their customers for global analysis of networks. This allows them to see trends across geographies, industries, and other variables. But some customers are uneasy with this aspect of AI/ML solutions. They don’t want even their anonymised data involved in this way.
The third biggest data-related challenge is network overhead. More than 36% of organisations are concerned about the network cost of moving vast data sets off-premises into a cloud-based data lake. This data transfer can sometimes consume too much bandwidth. Some vendors mitigate this issue by processing data at the network edge with local probes, which then forward metadata into the AI cloud for analysis. Organisations that are evaluating AI-driven networking solutions should ask potential vendors how they address this issue.
Finally, 32% of organisations told EMA that their data lacks granularity. They are unable to collect data at intervals that are short enough to provide their AI solutions with sufficient insight into their network. This issue can emerge in various ways. Some SD-WAN vendors limit the rates at which they collect network telemetry because the telemetry traffic can impact network performance.
Some monitoring tools limit intervals at which they poll networks with SNMP because higher polling rates can destabilise the monitoring platform. And some network switches and routers are limited in how often they can generate flow records due to performance impacts. More recently, some network vendors have started using switch silicon that is optimised to generate more granular data that could help mitigate this issue, but this hardware often comes at a premium cost.
Evaluate the network data you already gather
Even if you have no plans to adopt AI/ML solutions for network management, it’s always a good idea to review the state of network data in your organisation. Network operations teams often tell EMA that their biggest challenge in general is data quality, regardless of whether AI is on the horizon.
For instance, organisations should determine whether there are any blind spots in their network that may become glaringly obvious when AI starts looking at things. They should review the quality of the data their tools are collecting and retaining. Is this collected data prone to errors? The data should also adhere to standards. If tools are tagging data with metadata, will a third-party be able to parse it? Standardisation will ensure that it’s readable by other systems. Also, think about data-collection intervals. A lot can happen in the five or ten minutes between SNMP polling intervals.
(More about how AI/ML can optimise network operations is available at EMA’s free webinar about its latest research.)