This paper focuses on the recommendation of job ads to job seekers, exploiting proprietary data from the French Public Employment Service (PES) and focusing more specifically on low or unskilled workers. Besides the usual challenges of data sparsity, the signal to noise ratio is high (few job seekers have diplomas), and scalability requirements are paramount. As a first contribution, a two-tiered approach is designed to handle these requirements; its empirical validation shows significant computational gains with no performance loss compared to boosted tree ensembles representative of the state of the art. A second contribution is a methodology aimed to assess the impact of the non-stationarity of the item and user distributions. Specifically, during the last three periods (before, during and after the Covid lock-downs), the numbers of job ads and job seekers dramatically vary in some industries. A normalized recall indicator is proposed to filter out the impact of variations of the number of job ads. This normalization suggests that the same score function adapts to the multi-faceted changes of the environment, resulting in different recommendations but with similar accuracy as before - at least for the job seekers finding a job.