With Oracle migrating field that's FileField(null=True) to FileField(blank=True) changes column to NOT NULL (it shouldn't)

[Matt Hoskins]

Oracle stores empty strings as NULL values and Django has code to handle that, but there is a bug.

Background: I just tried using django-silk with an Oracle database as the back end and it was generating errors trying to safe a blank value in a FileField that had blank=True set. The error was: django.db.utils.IntegrityError: ORA-01400: cannot insert NULL into ("MISPORTAL"."SILK_REQUEST"."PROF_FILE")

As noted Django has code to deal with the fact Oracle stores empty strings as NULL values (and sets char fields as NULL rather than NOT NULL if blank=True), so it shouldn't happen. But looking through the migration history I saw that the schema originally defined the field with null=True and changed it to blank=True later, so there's an AlterField that comes into play where the null attribute on the field has changed.

From a quick skim of the code I believe Django has code to ignore migrations which change whether a column is nullable or not if connection.features.interprets_empty_strings_as_nulls is True but only if the field's get_internal_type() is "CharField" or "TextField" - that obviously doesn't include "FileField". So although Django will correctly create FileField columns where blank=True as "NULL" under Oracle, if they were originally null=True and are later migrated to just be blank=True then Django won't skip changing the column to "NOT NULL" and so you'll end up with the column being NOT NULL and empty strings in that field will result in the error as Oracle tries to save them as NULL values.

The bit which I think needs changing to fix this is in db/backends/base/schema.py and is in _alter_column_null_sql - I think just adding "FileField" to the list should do it. Having said that - I don't know enough about Django's internals but it does seem a little strange that the test in _alter_column_null_sql doesn't mirror the test that is in _iter_column_sql of checking "(field.empty_strings_allowed and not field.primary_key and self.connection.features.interprets_empty_strings_as_nulls)" rather than the specific type - although perhaps the test in _iter_column_sql got changed in the past and _alter_column_null_sql got missed as being altered to match it?