Let $X\rightarrow S$ be an arithmetic surface (this means, in the language of Qing Liu's book, $X$ is a regular integral scheme of dimension 2, projective and flat over a Dedekind scheme $S$ of dimension 1). Let $\eta$ be the generic point of $S$.
Let $\omega_{X/S}$ be the dualizing sheaf of $X/S$, which is invertible, hence there is a Cartier divisor $K_{X/S}$ such that $\omega_{X/S}\cong\mathcal{O}_X(K_{X/S})$.
At first, it seemed natural to me that since $\omega_{X/S}$ is a "relative object", $K_{X/S}$ should be linearly equivalent to a purely horizontal effective divisor - ie, is a nonnegative linear combination of horizontal prime divisors. However, this cannot always be the case, since if $X_\eta$ is genus $\ge 1$, then $K_{X/S}$ can be chosen to be effective, and if $X$ is not minimal (ie, contains an exceptional (vertical) divisor $E$), then $K_{X/S}\cdot E < 0$, which can only happen if $E$ is a component of $K_{X/S}$ (for any effective canonical divisor $K_{X/S}$).
Thus, the question is: If $X_\eta$ is say smooth of genus $\ge 1$ (or $\ge 2$ if $g = 1$ is special), and $X$ is minimal, then can we always find an effective canonical divisor $K_{X/S}$ which is a linear combination of horizontal prime divisors?
Is there any intuition for why this is/isn't/should/shouldn't be the case?