The recent increase in volume and complexity of available astronomical data\nhas led to a wide use of supervised machine learning techniques. Active\nlearning strategies have been proposed as an alternative to optimize the\ndistribution of scarce labeling resources. However, due to the specific\nconditions in which labels can be acquired, fundamental assumptions, such as\nsample representativeness and labeling cost stability cannot be fulfilled. The\nRecommendation System for Spectroscopic follow-up (RESSPECT) project aims to\nenable the construction of optimized training samples for the Rubin Observatory\nLegacy Survey of Space and Time (LSST), taking into account a realistic\ndescription of the astronomical data environment. In this work, we test the\nrobustness of active learning techniques in a realistic simulated astronomical\ndata scenario. Our experiment takes into account the evolution of training and\npool samples, different costs per object, and two different sources of budget.\nResults show that traditional active learning strategies significantly\noutperform random sampling. Nevertheless, more complex batch strategies are not\nable to significantly overcome simple uncertainty sampling techniques. Our\nfindings illustrate three important points: 1) active learning strategies are a\npowerful tool to optimize the label-acquisition task in astronomy, 2) for\nupcoming large surveys like LSST, such techniques allow us to tailor the\nconstruction of the training sample for the first day of the survey, and 3) the\npeculiar data environment related to the detection of astronomical transients\nis a fertile ground that calls for the development of tailored machine learning\nalgorithms.\n