Filling in missing values.
NaN at all, so missing data has to be resolved before anything else in a pipeline. These transformers learn what to fill from the training data and apply the same values later, which is what keeps a test split from influencing its own imputation.
Impute before scaling: a scaler fitted on data still containing NaN propagates it into the mean and variance. The "standard" and "full" pipeline presets already order the two correctly.
Missingness is sometimes informative — a blank field can mean "not applicable" rather than "unknown". Where that is true, record it as its own feature before imputing, or the signal is erased.