Then, what is wrong? The spike reflects the data. Rather, graphs with short time intervals failed to catch the spikes.
What actually happend is that after a slight gap, sharp decrease of network.in/out occured. It is possible something happend then, but it is beyonde this graph.
Anyway, as for the default strategy against missing values of original series for differences Formula, missing value: last could be safer. You may issue a feature request here, if you want.