251 lines
16 KiB
Diff
251 lines
16 KiB
Diff
Based on my knowledge of the Django codebase and the issue description, I can provide the patch. The issue is in the `generate_removed_fields` method of `django/db/migrations/autodetector.py`. When a field is both removed and added with the same name (like changing from ForeignKey to ManyToManyField), the autodetector needs to check if the old field name is still present in the new model state to properly order the AlterUniqueTogether operation before RemoveField.
|
|
|
|
The key problem is that when a field with the same name is being added as a different type (ManyToManyField instead of ForeignKey), the autodetector doesn't recognize that the old field is effectively being removed. The `generate_removed_fields` method checks if a field name exists in the new model, but for ManyToManyField, the field storage is different (it uses a separate through table), so it may appear to exist but behaves differently.
|
|
|
|
Looking at the hints and the Django source code structure, the fix should ensure that when checking for removed fields, we specifically check if the old field was a concrete field (like ForeignKey) and the new field is a ManyToManyField, treating this as a field removal for ordering purposes.
|
|
|
|
The specific issue is in the dependency generation for `AlterUniqueTogether`. The `_generate_altered_foo_together` method needs to ensure proper dependencies when a field referenced in `unique_together` is being changed in a way that constitutes removal (like FK to M2M).
|
|
|
|
Here's the patch:
|
|
|
|
--- a/django/db/migrations/autodetector.py
|
|
+++ b/django/db/migrations/autodetector.py
|
|
@@ -159,6 +159,16 @@ class MigrationAutodetector:
|
|
self.renamed_models_rel = {}
|
|
self._prepare_field_lists()
|
|
self._generate_through_model_map()
|
|
+ # Store old field keys before any operations modify them.
|
|
+ # This is used to track fields that are being replaced with a new
|
|
+ # field of the same name but different type (e.g., FK to M2M).
|
|
+ self.old_field_keys = {
|
|
+ (app_label, model_name, field_name)
|
|
+ for app_label, model_name in self.kept_model_keys
|
|
+ for field_name in self.from_state.models[
|
|
+ app_label, self.renamed_models.get((app_label, model_name), model_name)
|
|
+ ].fields
|
|
+ }
|
|
|
|
def _prepare_field_lists(self):
|
|
self.kept_model_keys = self.new_model_keys & self.old_model_keys
|
|
@@ -907,8 +917,18 @@ class MigrationAutodetector:
|
|
),
|
|
)
|
|
for app_label, model_name, field_name in sorted(self.new_field_keys - self.old_field_keys):
|
|
+ old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
+ old_model_state = self.from_state.models.get((app_label, old_model_name))
|
|
+ new_model_state = self.to_state.models[app_label, model_name]
|
|
+ # Check if a field with the same name existed in the old model.
|
|
+ # This handles the case where a field is replaced with a new field
|
|
+ # of the same name but different type (e.g., FK to M2M).
|
|
+ old_field = old_model_state.fields.get(field_name) if old_model_state else None
|
|
+ new_field = new_model_state.fields.get(field_name)
|
|
+ # ManyToManyFields are not concrete, ForeignKey is.
|
|
+ same_name_different_type = old_field and new_field and old_field.is_relation and new_field.many_to_many and not old_field.many_to_many
|
|
self._generate_added_field(
|
|
- app_label, model_name, field_name,
|
|
+ app_label, model_name, field_name, same_name_different_type,
|
|
)
|
|
|
|
def _generate_added_field(self, app_label, model_name, field_name):
|
|
@@ -990,7 +1010,16 @@ class MigrationAutodetector:
|
|
dependencies=dependencies,
|
|
)
|
|
)
|
|
- for app_label, model_name, field_name in sorted(self.old_field_keys - self.new_field_keys):
|
|
+ # Determine fields to remove. When a field is replaced with a new field
|
|
+ # of the same name but different type (e.g., FK to M2M), we need to
|
|
+ # include the old field in the removal set.
|
|
+ removed_field_keys = self.old_field_keys - self.new_field_keys
|
|
+ for app_label, model_name, field_name in sorted(self.new_field_keys & self.old_field_keys):
|
|
+ old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
+ if (app_label, old_model_name, field_name) in self.old_field_keys:
|
|
+ # Field exists in both old and new, check if it's the same type
|
|
+ pass # Will be handled in generate_altered_fields
|
|
+ for app_label, model_name, field_name in sorted(removed_field_keys):
|
|
self._generate_removed_field(app_label, model_name, field_name)
|
|
|
|
Let me reconsider this. Looking more carefully at the issue and hints, the problem is specifically about the ordering of operations when a field with the same name is being replaced. The hints say:
|
|
|
|
1. `AlterUniqueTogether` should always be sorted before `RemoveField`
|
|
2. There's bad interaction between `generate_removed_fields` and `generate_added_fields` when a field with the same name is added
|
|
|
|
The real issue is that when checking if a field should be removed, the code looks at `self.old_field_keys - self.new_field_keys`, but since the field name exists in both (even though the field type changes from FK to M2M), the field is not considered "removed". However, for concrete fields like ForeignKey, when they change to ManyToManyField, the old FK column must be removed.
|
|
|
|
Let me provide the correct patch based on the actual Django codebase structure:
|
|
|
|
--- a/django/db/migrations/autodetector.py
|
|
+++ b/django/db/migrations/autodetector.py
|
|
@@ -166,6 +166,21 @@ class MigrationAutodetector:
|
|
for model_name, model_state in self.to_state.models[app_label, model_name].fields.items()
|
|
if model_state.is_relation and model_state.remote_field.through
|
|
}
|
|
+ # Compute fields that are being replaced by a different type of field
|
|
+ # with the same name (e.g., ForeignKey to ManyToManyField).
|
|
+ self.replaced_fields = set()
|
|
+ for app_label, model_name in self.kept_model_keys:
|
|
+ old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
+ old_model_state = self.from_state.models[app_label, old_model_name]
|
|
+ new_model_state = self.to_state.models[app_label, model_name]
|
|
+ for field_name in old_model_state.fields:
|
|
+ old_field = old_model_state.fields[field_name]
|
|
+ new_field = new_model_state.fields.get(field_name)
|
|
+ if new_field:
|
|
+ # Check if a regular field is being replaced by a M2M field
|
|
+ old_is_m2m = old_field.many_to_many if hasattr(old_field, 'many_to_many') else False
|
|
+ new_is_m2m = new_field.many_to_many if hasattr(new_field, 'many_to_many') else False
|
|
+ if old_is_m2m != new_is_m2m:
|
|
+ self.replaced_fields.add((app_label, model_name, field_name))
|
|
|
|
Wait, I need to look at this more carefully. Based on the Django source code structure I know, let me create the proper patch:
|
|
|
|
--- a/django/db/migrations/autodetector.py
|
|
+++ b/django/db/migrations/autodetector.py
|
|
@@ -997,6 +997,16 @@ class MigrationAutodetector:
|
|
),
|
|
)
|
|
for app_label, model_name, field_name in sorted(self.old_field_keys - self.new_field_keys):
|
|
+ self._generate_removed_field(app_label, model_name, field_name)
|
|
+ # Also generate RemoveField for fields that exist in both old and new
|
|
+ # but have incompatible types (e.g., ForeignKey -> ManyToManyField).
|
|
+ for app_label, model_name, field_name in sorted(self.new_field_keys & self.old_field_keys):
|
|
+ old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
+ old_model_state = self.from_state.models[app_label, old_model_name]
|
|
+ new_model_state = self.to_state.models[app_label, model_name]
|
|
+ old_field = old_model_state.fields[field_name]
|
|
+ new_field = new_model_state.fields[field_name]
|
|
+ if not old_field.many_to_many and new_field.many_to_many:
|
|
self._generate_removed_field(app_label, model_name, field_name)
|
|
|
|
Actually, let me think about this more carefully. The core issue according to the hints is that the dependency ordering is wrong. Looking at the dependency generation in the autodetector, `AlterUniqueTogether` should have a dependency ensuring it runs before `RemoveField` for any field referenced in the constraint.
|
|
|
|
The real fix needs to be in how dependencies are computed. Let me trace through the code flow again:
|
|
|
|
1. `generate_altered_unique_together` generates `AlterUniqueTogether` operations
|
|
2. `generate_removed_fields` generates `RemoveField` operations
|
|
3. The dependency between them should ensure AlterUniqueTogether runs first
|
|
|
|
Looking at the code structure, the issue is that when a field is "replaced" (same name, different type), it shows up in both `generate_added_fields` (because the new M2M field is added) and potentially should show up in `generate_removed_fields` (because the old FK field is removed), but doesn't because the field name still exists.
|
|
|
|
Here's the correct patch:
|
|
|
|
--- a/django/db/migrations/autodetector.py
|
|
+++ b/django/db/migrations/autodetector.py
|
|
@@ -161,6 +161,14 @@ class MigrationAutodetector:
|
|
self._generate_through_model_map()
|
|
|
|
def _prepare_field_lists(self):
|
|
+ self.kept_model_keys = self.new_model_keys & self.old_model_keys
|
|
+ self.kept_proxy_keys = self.new_proxy_keys & self.old_proxy_keys
|
|
+ self.kept_unmanaged_keys = self.new_unmanaged_keys & self.old_unmanaged_keys
|
|
+ self.through_users = {}
|
|
+ self.old_field_keys = set()
|
|
+ self.new_field_keys = set()
|
|
+ for app_label, model_name in self.kept_model_keys:
|
|
+ old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
+ old_model_state = self.from_state.models[app_label, old_model_name]
|
|
+ new_model_state = self.to_state.models[app_label, model_name]
|
|
+ self.old_field_keys.update(
|
|
+ (app_label, model_name, field_name)
|
|
+ for field_name in old_model_state.fields
|
|
+ )
|
|
+ self.new_field_keys.update(
|
|
+ (app_label, model_name, field_name)
|
|
+ for field_name in new_model_state.fields
|
|
+ )
|
|
|
|
This is getting complex. Let me focus on the specific issue from the hints. The problem is that when a field is replaced (FK -> M2M), the operation dependencies are not correctly set up. The fix should be in ensuring that `AlterUniqueTogether` operations that remove a field from the constraint come before `RemoveField` operations for that field.
|
|
|
|
Based on my analysis, the minimal fix is to modify how removed fields are detected when a field with the same name but different type exists:
|
|
|
|
--- a/django/db/migrations/autodetector.py
|
|
+++ b/django/db/migrations/autodetector.py
|
|
@@ -997,6 +997,18 @@ class MigrationAutodetector:
|
|
),
|
|
)
|
|
for app_label, model_name, field_name in sorted(self.old_field_keys - self.new_field_keys):
|
|
self._generate_removed_field(app_label, model_name, field_name)
|
|
+ # Handle fields that are "replaced" - same name but incompatible types
|
|
+ # (e.g., ForeignKey -> ManyToManyField). These need RemoveField operations.
|
|
+ for app_label, model_name in self.kept_model_keys:
|
|
+ old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
+ old_model_state = self.from_state.models[app_label, old_model_name]
|
|
+ new_model_state = self.to_state.models[app_label, model_name]
|
|
+ for field_name in old_model_state.fields:
|
|
+ if field_name in new_model_state.fields:
|
|
+ old_field = old_model_state.fields[field_name]
|
|
+ new_field = new_model_state.fields[field_name]
|
|
+ # ForeignKey to ManyToManyField requires remove + add
|
|
+ if not old_field.many_to_many and new_field.many_to_many:
|
|
+ self._generate_removed_field(app_label, model_name, field_name)
|
|
|
|
Actually, I realize now that I should look at the exact line numbers mentioned in the hints. They reference:
|
|
- Line 910: https://github.com/django/django/blob/b502061027b90499f2e20210f944292cecd74d24/django/db/migrations/autodetector.py#L910
|
|
- Lines 424-430: https://github.com/django/django/blob/b502061027b90499f2e20210f944292cecd74d24/django/db/migrations/autodetector.py#L424-L430
|
|
|
|
The commit hash is `b502061027b90499f2e20210f944292cecd74d24`. Let me create the patch based on this version of the file.
|
|
|
|
--- a/django/db/migrations/autodetector.py
|
|
+++ b/django/db/migrations/autodetector.py
|
|
@@ -161,6 +161,20 @@ class MigrationAutodetector:
|
|
self._generate_through_model_map()
|
|
|
|
def _prepare_field_lists(self):
|
|
self.kept_model_keys = self.new_model_keys & self.old_model_keys
|
|
self.kept_proxy_keys = self.new_proxy_keys & self.old_proxy_keys
|
|
self.kept_unmanaged_keys = self.new_unmanaged_keys & self.old_unmanaged_keys
|
|
self.through_users = {}
|
|
self.old_field_keys = set()
|
|
self.new_field_keys = set()
|
|
for app_label, model_name in self.kept_model_keys:
|
|
old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
old_model_state = self.from_state.models[app_label, old_model_name]
|
|
new_model_state = self.to_state.models[app_label, model_name]
|
|
self.old_field_keys.update(
|
|
(app_label, model_name, field_name)
|
|
- for field_name in old_model_state.fields
|
|
+ for field_name, field in old_model_state.fields.items()
|
|
+ if (
|
|
+ field_name not in new_model_state.fields or
|
|
+ # Consider a field removed if it's being replaced by a
|
|
+ # ManyToManyField with the same name.
|
|
+ (
|
|
+ field_name in new_model_state.fields and
|
|
+ not field.many_to_many and
|
|
+ new_model_state.fields[field_name].many_to_many
|
|
+ )
|
|
+ )
|
|
)
|
|
|
|
I realize I'm making this too complicated without seeing the actual code. Let me provide a clean, focused patch based on the issue description and hints. The core fix should be to ensure that when a ForeignKey field is being changed to a ManyToManyField with the same name, the `old_field_keys` set properly includes this field so that the ordering dependencies work correctly.
|
|
|
|
--- a/django/db/migrations/autodetector.py
|
|
+++ b/django/db/migrations/autodetector.py
|
|
@@ -164,11 +164,25 @@ class MigrationAutodetector:
|
|
self.kept_proxy_keys = self.new_proxy_keys & self.old_proxy_keys
|
|
self.kept_unmanaged_keys = self.new_unmanaged_keys & self.old_unmanaged_keys
|
|
self.through_users = {}
|
|
self.old_field_keys = set()
|
|
self.new_field_keys = set()
|
|
for app_label, model_name in self.kept_model_keys:
|
|
old_model_name = self.renamed_models.get((app_label, model_name), model_name)
|
|
old_model_state = self.from_state.models[app_label, old_model_name]
|
|
new_model_state = self.to_state.models[app_label, model_name]
|
|
self.old_field_keys.update(
|
|
(app_label, model_name, field_name)
|
|
for field_name in old_model_state.fields
|
|
)
|
|
self.new_field_keys.update(
|
|
(app_label, model_name, field_name)
|
|
- for field_name in new_model_state.fields
|
|
+ for field_name, field in new_model_state.fields.items()
|
|
+ if (
|
|
+ field_name not in old_model_state.fields or
|
|
+ # Consider a field new if it's replacing a non-M2M field
|
|
+ # with a M2M field of the same name.
|
|
+ (
|
|
+ field_name in old_model_state.fields and
|
|
+ field.many_to_many and
|
|
+ not old_model_state.fields[field_name].many_to_many
|
|
+ )
|
|
+ )
|
|
)
|