Surrogate models replace computationally expensive simulations of physically-based models to obtain accurate results at a fraction of the time. These surrogate models, also known as metamodels, have been employed for analysis, control, and optimisation of water distribution and urban drainage systems. With the advent of machine learning (ML), water engineers have increasingly resorted to these data-driven techniques to develop metamodels of urban water networks. In this manuscript, we review 31 recent papers on ML-based metamodeling of urban water networks to outline the state-of-the-art of the field, identify outstanding gaps, and propose future research directions. For each paper, we critically examined the purpose of the metamodel, the metamodel characteristics, and the applied case study. The review shows that current metamodels suffer several drawbacks, including i) the curse of dimensionality, hindering implementation for large case studies; ii) black-box deterministic nature, limiting explainability and applicability; and iii) rigid architecture, preventing generalization across multiple case studies. We argue that researchers should tackle these issues by resorting to recent advancements in ML concerning inductive biases, robustness, and transferability. The recently developed Graph Neural Network architecture, which extends deep learning methods to graph data structures, is a preferred candidate for advancing surrogate modelling in urban water networks. Furthermore, we foresee increasing efforts for complex applications where metamodels may play a fundamental role, such as uncertainty analysis and multi-objective optimisation. Lastly, the development and comparison of ML-based metamodel can benefit from the availability of new benchmark datasets for urban drainage systems and realistic complex networks.