Simply running twice cells 13 & 14 of the notebook that read and plot point single variable time-series for a point will reproduce this issue where the first run will have the valid values but second will be all nans. I encountered this when parallelizing reading of the files with dask that results in multiple reads and the unexpected result.