Some bug fixes and future updates will lead to unavoidable changes that are non-compatible with older software versions. The code needs to be revisited to make sure that such backward-compatibility issues are robustly detected and errors raised. The models already keep track of software version used for training of the model, but I don't think (need to check this) the current code checks against specific versions yet.
A common strategy needs to be defined to account for such compatibility issues. One option could be to check in those places of the code in which in-compatible code is introduced in certain software versions. In any case, the goal should be to keep trained models available - independent of the software version - for as long as its possible, i.e. no incompatible changes are introduced.
Some bug fixes and future updates will lead to unavoidable changes that are non-compatible with older software versions. The code needs to be revisited to make sure that such backward-compatibility issues are robustly detected and errors raised. The models already keep track of software version used for training of the model, but I don't think (need to check this) the current code checks against specific versions yet. A common strategy needs to be defined to account for such compatibility issues. One option could be to check in those places of the code in which in-compatible code is introduced in certain software versions. In any case, the goal should be to keep trained models available - independent of the software version - for as long as its possible, i.e. no incompatible changes are introduced.