Machine learning is an inescapable buzzword for many in the operations sector. Even friends and colleagues tend to make us aware of a new ML tool that may or may not be useful. While there are many ML tools in the market, not all are suitable for every business. Some tools, when tested, struggle to solve basic, everyday use cases. Therefore, when evaluating ML tools, other deeper questions and issues do arise. Let us analyze these questions and see if we can get a better idea of what to expect when evaluating AI/ML tools.
Why do I need to consider Machine Learning tools?
The operations side of any IT business typically involves an endless amount of firefighting. The errors encountered during deployment of a simple new patch or upgrade can have a series of ramifications and process delays. Not to mention, it can eat up a large chunk of your own invaluable time!
While tools can be configured to identify and resolve repetitive issues, new issues are a whole different ball game. As business applications, models, and processes evolve, even an experienced operations expert may find it difficult to predict when and how things can fail. Uptime included in SLAs tend to be affected when new issues crop up and require complex solutions.
The difference between a good and bad ML tool is its ability to help in this exact scenario. A good ML tool can help find different variations, usage patterns, detect anomalies and much more! Swift solutions come about thanks to the ability of the tool to get to the heart of the problem rather than just analyze symptoms.
What do I need to look for in Machine Learning tools?
An ML tool that can easily be tailored is the solution you need. One-size-fits-all solutions simply don’t work. The data and requirements vary wildly from sector to sector. A stock market related machine learning algorithm cannot solve infrastructure and capacity forecasting issues. A psychometric analytical program cannot predict application outages by looking at patterns. Each business problem needs specialized algorithms and techniques to understand and solve the problem.
Data Scientist use case
We work with data scientists on a regular basis. Points of differentiation and the ability of the tool to understand the difference between univariate analysis and multivariate analysis are crucial.
As the name suggests, an univariate analysis involves us looking for a single variable and analyzing its distribution across multiple points. Data scientists typically use one of these 3 techniques:
- Summary Statistics
- Frequency Distribution
- Charts
An univariate analysis helps find the optimum range or threshold of any metric. If you need to monitor CPU utilization, you can set rules to ensure that the CPU utilization stays between 20-78% between 9 AM and 9 PM on a specific host.
Multivariate analysis on the other hand analyses two or more variables across data points. Three commonly used techniques are:
- Scatter Plots
- Correlation Coefficients
- Simple Linear Regressions
Typically, multivariate analysis is used to find dependencies and correlations across multiple metrics. Keeping the same CPU utilization use case in mind, we can now analyze data during concurrent sessions and set optimal thresholds based on a number of dependencies determined by our multivariate analysis.
In reality, CPUs can carry out an univariate analysis easily but can struggle to execute multivariate analysis for problem solving. Hardware is a limiting factor and utilizing a GPU instead of a CPU is a great alternative to ensure speed is never compromised.
Most APM and AIOps tools in the market only use univariate analysis or simulate a multivariate analysis using a time-series variable.
The ML tool you choose for your business can depend on your need for univariate or multivariate analysis and your company’s ability to invest in GPU-based hardware.
What is ML Ops?
The role of a data scientist is quite limited when it comes to providing techniques for problem solving using AI/ML. Off-the-shelf ML Ops solutions, coupled with excellent data science processes, are the way to go! ML Ops teams run development pipelines to enable data cleansing, create models, enable feedback looks etc.
Summary
For new age applications, the best approach, without a doubt, involves using AI/ML to solve business problems. Dynamism is commonplace and investing in a good ML Ops team is key for continuous success.