To better align business and IT objectives, enterprise organizations should focus on the core "problems" that individual business units face today in driving out real consumer value. Until the roadblocks and inhibitors — and, ultimately, the resultant technical debt — are removed from the equation, large enterprise organizations will continue struggling to succeed ...
Data science and machine learning algorithms have become pervasive throughout the modern consumer world. There are many successful applications of machine learning in consumer products that we use on a daily basis, including:
■ Movie, music and product recommendations
■ Ad targeting
■ Web search
But, when we look at the Ops world, we find that there is no breakthrough product that incorporates these same machine learning innovations.
A relevant comparison is with page-level and host-level features (for example, page rank for URLs or host rank for hosts) used in search ranking. These features are typically a function of the WebMap (the massive graph where nodes are URLs, and edges are hyperlinks between URLs). The page rank algorithm allows the ranking of URLs in the WebMap based on the hyperlinks between them. It is a very effective way to get a reasonable estimate of the overall importance of a URL.
What if we used similar ideas to rank the hundreds or sometimes thousands of alerts that operations engineers receive, especially when they are managing hundreds of machines? What is the equivalent of the WebMap in the Ops world?
Another relevant example is provided by duplicate web page detection. These algorithms run as MapReduce jobs on massive Hadoop clusters (thousands of machines) and detect duplicate pages across tens of billions of web pages. When the mappers or reducers fail or when there are performance degradations, hundreds of alerts are generated, many of them for the same underlying root cause.
What if we applied the techniques of web page duplicate detection to eliminate the duplicate and unnecessary alerts received by Ops?
A third big challenge is personalization of content. Personalization is a well-studied problem in the consumer space, with user feedback — both implicit (clicks and actions) and explicit (reviews and ratings) — contributing critical inputs to the learning algorithms. Employing this type of machine learning means that the more time a user spends with a product, the better their user experience will be.
What if we incorporated feedback to learn Ops users’ preferences and continuously improve the accuracy of alert generation and alert ranking?
The answers to these questions will become evident as we bring the innovations in data science and machine learning that are commonplace in the consumer world to the Ops world. DevOps teams need, in effect, an “expert assistant” that can learn their application and system environment, detect and correlate failures, and make recommendations that drive increased focus and productivity — even as everything is continuously changing. It’s time for Ops to get smarter.
Amit Sasturkar is Co-Founder and CTO of OpsClarity.