The algorithm for retrieving observed data for assimilation, especially in cases of nudging at the beginning of forecast runs, goes too much back (and forward) in time and retrieves a lot of useless observations. The interval should be reduced somehow.