Apparently we get some overhead when mlr3pipelines builds tasks with many BackendCbinds. One way to fix this would be if there were an option to "flatten" cbinded tasks.
Suggested interface:

Task$flatten(force = FALSE)  # default

creates a task with a single BackendDataTable, unless this is for some reason a bad idea, e.g. when a backend is a database backend. (A Backend class would need to report whether flattening is a "bad idea", possibly with an active binding, e.g. a database backend could say flattening is okay if the size is less than X MB)

Setting force = TRUE should OTOH flatten the task always, equivalent to creating a new task with the task$data().

Example: TaskClassif that consists of two cbinded data.tables that were cbinded with a database backend:
(abbreviating (DataBackend as DB)

                TaskClassif
                  |
               DBCbind
              /       \
         DBCbind      DBDataBase
        /       \ 
 DBDataTable DBDataTable

$flatten(force = FALSE):

                TaskClassif
                  |
               DBCbind
              /       \
     DBDataTable      DBDataBase

$flatten(force = TRUE):

               TaskClassif
                  |
               DBDataTable

We could think whether it is a good idea if mlr3pipelines does this with all its output tasks by default.

Another question is whether that should be an in-place operation that swaps out a task's data backend, or whether this should create a new task.

Another question is what to do with columns that do not have any column role. Maybe a good default would be to drop backends that do not provide columns that have a role (and are therefore ignored in many cases).

Maybe we would want to have a DataBackendMultiCBind that can cbind multiple sources, so even a task that has many different database backends will only be one level deep at the most after flattening. The $flatten(force = FALSE) -operation would have to check, for each column, if it comes from a data backend that reports it does not want to be flattened. There should be a method in DataBackend that does this recursively. $flatten() would then construct the desired DataBackendMultiCBind.