Artificial intelligence (AI) and large language models (LLMs) have dominated the tech scene over the past year. As a byproduct, vendors in nearly every tech sector are adding AI capabilities and scrambling to promote how their products and services use it. 

This trend has also made its way to the observability and monitoring space. However, the AI solutions coming to market often feel like putting a square peg in a round hole. While AI can significantly impact certain areas of observability, it is not a fit for others. In this article, I’ll share my views on how AI can and cannot support an observability practice – at least right now.

The Long Tail of Errors

The very nature of observability makes ‘prediction’ in the traditional sense unfeasible. In life, certain ‘act of God’ types of events can impact business and are impossible to predict – weather-related events, geopolitical conflicts, pandemics, and more. These events are so rare and capricious that it’s implausible to train an AI model to predict when one is imminent.

The long tail of potential errors in application development mirrors this. In observability, many errors may happen only once, such that you may never see them happen again in your lifetime, while other types of errors may occur daily. So, if you’re looking to train a model that will completely understand and predict all the ways things could go wrong in an application development context, you’re likely to be disappointed.

Poor Quality Data

Another way that AI needs to improve in observability is its inability to make a distinction between details that are irrelevant, and those that are not. In other words, AI can pick up on small, inconsequential aberrations with a big impact on your results.

For example, previously, I worked with a customer training an AI model with hours of basketball footage to predict successful versus unsuccessful baskets. There was one big issue: all footage of an unsuccessful basket included a timestamp on the video. So, the model determined timestamps have an impact on the success of a shot (not the result we were looking for).

Observability practices often work with imperfect data – unneeded log contents, noisy data, etc. When you introduce AI without cleaning up this data, you create the possibility of false positives – as the saying goes, “garbage in and garbage out.” Ultimately, this can leave organizations in a more vulnerable position of alert fatigue.

Where AI Does Fit Observability

So, where should we be using AI in observability? One area where AI can add a lot of value is in baselining datasets and detecting anomalies. In fact, many teams have been using AI for anomaly detection for quite some time. In this use case, AI systems can, for example, understand what “normal” activity is across different seasonalities and flag when it detects an outlier. In this way, AI can give teams a proactive heads-up when something may be going awry.

Another area where AI can be helpful is by shortening the learning curve when adopting a new query language. Several vendors are currently working on natural language query translators driven by AI. A natural language translator is an excellent way to lower the entry barriers when using a new tool. It frees up practitioners to focus on the flow and the practice itself rather than the pipes, semicolons, and all other nuances that come with learning a new syntax.

What to Focus on Instead

Whether beginning a journey with AI or making any other improvement, understanding usage trends is essential to optimizing the value of an observability practice. Improving a system without understanding its usage is akin to throwing darts in a pitch-black room. If no one uses the observability system, it’s pointless to have it. Many different analytics can help you know who’s using the system and, conversely, who isn’t using the system that should be.

Practitioners should focus on usage related to the following:

  • User-generated content – are users creating alerts or dashboards? How often are they being viewed? How delayed is the data getting to these dashboards, and can this be improved?
  • Queries – how often are you running queries powering dashboards and alerts?  Are queries fast or slow, and could they be optimized for performance? Understanding and improving query speed can improve development velocity for core functions.
  • Data – what volume is stored, and from what sources? How much of the stored data is actually queried?  What are the hotspots/dead zones, and can storage be tiered in a manner so as to optimize cloud storage costs?

Closing Thoughts

I believe that AI is currently at the peak of the hype curve. In an application development setting, pretending AI does what it doesn’t do – i.e., predict root causes and recommend specific remediations – is not going to propel us to the part after all the hype when the technology actually gets useful. There are very real ways that AI can turn the gears on observability improvements today – and this is where we should be focused.