Adding tables or table columns to a database schema is comparatively straightforward. The new table or column does not collide with older application versions, since they are unknown to them.

New tables that resulted from new domain entities can simply be added to the schema, resulting in version N+1.

New table columns that define certain constraints, such as not null or unique, need to take care of the current state of the table. The old application version can still write to the table; it will ignore the new column. Therefore, constraints can not necessarily be satisfied. New columns first need to be nullable and without further constraints. The new application version has to deal with empty values in that column, presumably null values, which originate from the old application version.

Only the next version (N+2) will then, after the current deployment has been completed, contain the correct constraints. This means that adding a column that defines constraints needs at least two separate deployments. The first deployment adds the column and enhances the application's model in a null-safe way. The second deployment makes sure all contained values fulfill the column constraints, adds the constraints, and removes the null-safe behavior. These steps are, of course, only required, if the column target state defines constraints.

Rollbacks to the old versions work in a similar way. Rolling back to the intermediate deployment (N+2 to N+1) requires the constraints to be removed again.

Rolling back to the original state (N+0) would remove the whole column. However, data migrations should not remove data that is not transferred somewhere else. Rolling back to the state without the column could also simply leave the column untouched so as not to lose data. The question the business experts have to answer is: What happens with the data that was added in the meantime? Intentionally not deleting this data could be a reasonable approach. However, when the column is added again, the rollout script needs to take already existing columns into consideration.