geom_bar: aes 'fill' and trans 'log10' break y scale

[ASmirnov-HORIS]

Example:

n = 1000
np.random.seed(42)
data = {
    "x": np.random.choice(["a", "b", "c"], p=[.49, .49, .02], size=n),
    "g": np.random.choice(["X", "Y"], size=n),
}
ggplot(data, aes(x="x", fill="g")) + geom_bar() + scale_y_continuous(trans='log10')

Output:

Label values on y axis are much greater than expected. There is no such trouble if fill="g" or trans='log10' will be removed.

[alshan]

Artem Smirnov (@ASmirnov-HORIS) is this still the observed behavior in 4.3.0 ?

[ASmirnov-HORIS]

Yes. Now only powers of 10 are left on the y-axis, but the values are still larger than they should be.

[alshan]

The thing is, in case of 'stack' position adj. we sum transformed values. And if the transform is 'log' transform we have:
log(256) + log(236) ~= 4.78 (and "unlog": 10^4.78 ~= 60255)

instead of expected:
log(256 + 236) ~= 2.69