Infrastructure-as-Code (IaC) is an emerging software engineering practice that leverages source code to facilitate automated configuration of software systems’ infrastructure. Like other software artifacts, severe defects could occur in IaC manifests, which can result in breaking millions of online services. To help developers early identify and fix IaC defects, several research efforts have introduced IaC defect prediction models at the file level. However, the granularity of the proposed approaches remains relatively coarse. Thus, identifying IaC bugs is a non-trivial task since IaC files may have many lines to inspect. To alleviate this issue, we aim to predict IaC defects at the fine-grained level by focusing on IaC blocks, which are code units that encapsulate specific behaviours within the IaC file. We evaluated our approach by conducting an empirical study on 19 open-source Terraform-based projects. We applied 6 machine learning algorithms to study to what extent they can predict defective blocks based on a mixture of code, process, and change level metrics. The results indicate that the LightGBM model emerged as the best model, achieving an average of 0.21 in the MCC and 0.71 in the AUC scores. We found that the developer’s experience and the relative number of added lines tend to be the most important features. Additionally, we found that blocks belonging to the most frequent types are more defect-prone. Our defect prediction models have also shown sensitivity to concept drift. This indicates that IaC practitioners should regularly retrain the defect models. We also suggest that practitioners stay attentive to changes in the syntax of block attributes, as these tend to be prone to modifications over time by IaC tool builders.