Data-Driven Methods for Data Center Operations Support

During the last decade, cloud technologies have been evolving at an impressive pace, such that we are now living in a cloud-native era where developers can leverage on an unprecedented landscape of (possibly managed) services for orchestration, compute, storage, load-balancing, monitoring, etc. The possibility to have on-demand access to a diverse set of configurable virtualized resources allows for building more elastic, flexible and highly-resilient distributed applications. Behind the scenes, cloud providers sustain the heavy burden of maintaining the underlying infrastructures, consisting in large-scale distributed systems, partitioned and replicated among many geographically dislocated data centers to guarantee scalability, robustness to failures, high availability and low latency. The larger the scale, the more cloud providers have to deal with complex interactions among the various components, such that monitoring, diagnosing and troubleshooting issues become incredibly daunting tasks. To keep up with these challenges, development and operations practices have undergone significant transformations, especially in terms of improving the automations that make releasing new software, and responding to unforeseen issues, faster and sustainable at scale. The resulting paradigm is nowadays referred to as DevOps. However, while such automations can be very sophisticated, traditional DevOps practices fundamentally rely on reactive mechanisms, that typically require careful manual tuning and supervision from human experts. To minimize the risk of outages—and the related costs—it is crucial to provide DevOps teams with suitable tools that can enable a proactive approach to data center operations. This work presents a comprehensive data-driven framework to address the most relevant problems that can be experienced in large-scale distributed cloud infrastructures. These environments are indeed characterized by a very large availability of diverse data, collected at each level of the stack, such as: time-series (e.g., physical host measurements, virtual machine or container metrics, networking components logs, application KPIs); graphs (e.g., network topologies, fault graphs reporting dependencies among hardware and software components, performance issues propagation networks); and text (e.g., source code, system logs, version control system history, code review feedbacks). Such data are also typically updated with relatively high frequency, and subject to distribution drifts caused by continuous configuration changes to the underlying infrastructure. In such a highly dynamic scenario, traditional model-driven approaches alone may be inadequate at capturing the complexity of the interactions among system components. DevOps teams would certainly benefit from having robust data-driven methods to support their decisions based on historical information. For instance, effective anomaly detection capabilities may also help in conducting more precise and efficient root-cause analysis. Also, leveraging on accurate forecasting and intelligent control strategies would improve resource management. Given their ability to deal with high-dimensional, complex data, Deep Learning-based methods are the most straightforward option for the realization of the aforementioned support tools. On the other hand, because of their complexity, this kind of models often requires huge processing power, and suitable hardware, to be operated effectively at scale. These aspects must be carefully addressed when applying such methods in the context of data center operations. Automated operations approaches must be dependable and cost-efficient, not to degrade the services they are built to improve. iv