When using col_wrap, seaborn builds the axes manually using fig.add_subplot (see:

This breaks the use of col/row for share{x,y} because add_subplot references a base axis, which is always set to axes_0.
Changing the reference axes to one on each row/column fixes this.

Example:

import seaborn as sns
tips = sns.load_dataset("tips")
tips.loc[tips['size']==6, 'tip'] += 100 # create a obvious skew in data
g = sns.FacetGrid(tips, col="size", col_wrap=3, sharey='row')
g.map_dataframe(sns.scatterplot, x="total_bill", y="tip")

The current implementation produces:

Suggested behavior:

Note that the first row is unaffected by the skewed data in the second row.

I have a working patch in my fork that I can submit as a PR, if this behavior makes sense.