Handle transaction rollbacks

[lukebryant]

Currently, if we do:

tm.field = 'new field value'
tm.get_dirty_fields() #returns {'field': 'original field value'}
try:
    with transaction.atomic(): 
        tm.save()
        raise Exception #force a rollback
except Exception:
     pass
tm.get_dirty_fields() #returns {}

This behaviour is problematic, because field in-memory and database values are still different.

It seems like this will be tricky to solve.

We could have a property uncommitted_dirty_fields, which we could clear in an on_commit() signal, but I can't think of a way of setting dirty_fields = uncommitted_dirty_fields in the case of a rollback, since there is no on_rollback() signal.

edit: If we had some way of uniquely identifying the current transaction (I can't see one in current django api), we might be able set dirty_fields = uncommitted_dirty_fields inside get_dirty_fields() if transaction_at_time_of_last_save != current_transaction. Then there'd be some additional complexity from rollbacks inside nested transactions.

[romgar]

Hey @lukebryant!
Interesting problem...
We could create a custom decorator / context manager that would wrap the transaction atomic and add this missing logic, but that's not really ideal as people would have to know that they should use this custom feature from dirtyfields instead of the default Django API...

[lukebryant]

Hey @lukebryant!
Interesting problem...
We could create a custom decorator / context manager that would wrap the transaction atomic and add this missing logic, but that's not really ideal as people would have to know that they should use this custom feature from dirtyfields instead of the default Django API...

I was thinking this wouldn't work because the context manager would have no way of getting references to all the models that are changed within its scope. Am I missing something, or did you mean that the context manager would take a list of potentially mutating models as arguments?

edit: thanks for quick response btw :)

[romgar]

Yeah you would need to pass these as arguments. Far from ideal for sure...

[romgar]

Right now, I would at least document the issue and ask people to manually save and restore dirty fields for their affected model in case of rollback.