On the limitations of Infrastructure as Code: some practical case studies

As of 2026, "Infrastructure as Code" (IaC) approaches have established theirselves as a new technique for infrastructure provisioning. One of the most popular IaC tools is Terraform, which allows to describe needed "data center based" infrastructure (e.g. fleets of virtual machines) through declarative configuration files. Abstracting over cloud providers' resource management web APIs in a standardized way and leveraging a declarative syntax, Terraform emerges as the "Holy Grail" of IaC. Terraform's focus is mostly on infrastructural objects rather than on what actually gets executed on them. For example, despite being able to describe virtual machines, Terraform can't fully manage configuration for their running software. This is, however, a voluntary limitation of the tool, which can be addressed by integrating Terraform with ad hoc solutions suitable for more "imperative like" operations. Cloud-init represents one of the aforementioned solutions and, in combination with Terraform, realizes a complete IaC system. In any case the resulting system still presents certain limitations and weaknesses, both considering its two components individually and the way they interact with one another. One method for bringing to light these faults is conducting "in-depth" experiments with both tools together, employing them under practical use cases in order to achieve a sufficient level of representativeness. From November 2025 to March 2026 I've been participating to an internship at Nucleus S.R.L., a network and IT consulting company based in Udine: during the experience I had the opportunity to perform the needed experiments in that sense. Both Terraform and cloud-init were leveraged, using versions 1.13 and 24.3 respectively. A wide percentage of their known functionalities was unleashed, especially considering the former tool, which gave me the opportunity to assess a series of interesting limitations and weaknesses of the "Terraform - cloud-init duo". The aim of this thesis is therefore to provide the reader with the achieved results: although deriving from domain specific use cases, such results may anyway serve as a base to highlight some IaC shortcomings which still have to be widely acknowledged.