Magnetic Resonance Fingerprinting (MRF) is theoretically more efficient than steady-state quantitative MRI techniques because it exploits dynamic behavior to enhance differences in signals obtained from tissues with different relaxation parameters. In practice, MRF often struggles to deliver the predicted performance, requiring careful adjustment of sequence parameters such as flip-angles, repetition times and k-space sampling patterns. MRF sequences result in a highly undersampled dynamic image series; state-of-the-art methods now exploit a temporal low-rank (TLR) reconstruction approach to help resolve some of the resulting undersampling artifacts. While successful, the TLR reconstruction mixes signals across space and through time, obscuring how sampling might be optimized for best results. This work explores optimal sampling for TLR reconstruction of MRF. We examine conditioning of the reconstruction problem as a predictor of image quality, and propose an effective optimization algorithm for k-space sampling on regular grids. Based on this, we compare different sampling schemes in simulations and real phantom experiments. We explain how undersampling generates aliasing, enhances noise and errors, and demonstrate how parallel-imaging improves this. We also conclude that the final reconstruction quality depends on a combination of undersampling, the expected signal distribution and errors present in real scans, and point towards how these might be included in further optimization efforts.