ABSTRACT We introduce SuperNNova, an open-source supernova photometric classification framework that leverages recent advances in deep neural networks. Our core algorithm is a recurrent neural network (RNN) that is trained to classify light curves using only photometric information. Additional information such as host-galaxy redshift can be incorporated to improve performance. We evaluate our framework using realistic supernova simulations that include survey detection. We show that our method, for the type Ia versus non-Ia supernova classification problem, reaches accuracies greater than 96.92 ± 0.09 without any redshift information and up to 99.55 ± 0.06 when redshift, either photometric or spectroscopic, is available. Further, we show that our method attains unprecedented performance for the classification of incomplete light curves, reaching accuracies >86.4 ± 0.1 (>93.5 ± 0.8) without host-galaxy redshift (with redshift information) 2 d before maximum light. In contrast with previous methods, there is no need for time-consuming feature engineering and we show that our method scales to very large data sets with a modest computing budget. In addition, we investigate often neglected pitfalls of machine learning algorithms. We show that commonly used algorithms suffer from poor calibration and overconfidence on out-of-distribution samples when applied to supernova data. We devise extensive tests to estimate the robustness of classifiers and cast the learning procedure under a Bayesian light, demonstrating a much better handling of uncertainties. We study the benefits of Bayesian RNNs for SN Ia cosmology. Our code is open sourced and available on github1.