spectacular_settings caching bug in enum overrides

[fledman]

Describe the bug
The load_enum_name_overrides function uses a decorator to cache function execution. However, the function internally uses the global variable spectacular_settings.ENUM_NAME_OVERRIDES, which is actually not constant. So whichever set of settings are loaded during the first execution of this function are implicitly baked into the cached result.

Specifically, if you are serving multiple schemas, each one can have its own custom settings. The way this works is that a context manager wraps the core view logic, patching the global settings object:

drf-spectacular/drf_spectacular/views.py

Lines 79 to 84 in 2501bfd

	with patched_settings(self.custom_settings):
	if settings.USE_I18N and request.GET.get('lang'):
	with translation.override(request.GET.get('lang')):
	return self._get_schema_response(request)
	else:
	return self._get_schema_response(request)

drf-spectacular/drf_spectacular/settings.py

Lines 279 to 283 in 2501bfd

	try:
	spectacular_settings.apply_patches(patches)
	yield
	finally:
	spectacular_settings.clear_patches()

The end result is that your expected enum overrides are not consistently applied, leading to unpredictable schema generation.

This bug is not new. It impacts both the historical implementation:

drf-spectacular/drf_spectacular/plumbing.py

Lines 862 to 865 in 9268276

	@cache
	def load_enum_name_overrides():
	overrides = {}
	for name, choices in spectacular_settings.ENUM_NAME_OVERRIDES.items():

as well as the recent changes to fix translation:

drf-spectacular/drf_spectacular/plumbing.py

Lines 863 to 870 in 2501bfd

	def load_enum_name_overrides():
	return _load_enum_name_overrides(get_language())
	@functools.lru_cache()
	def _load_enum_name_overrides(language: str):
	overrides = {}
	for name, choices in spectacular_settings.ENUM_NAME_OVERRIDES.items():

To Reproduce
Suppose your drf project has multiple installed applications. Each app can serve its own schema, with potentially unique drf-specatcular settings:

# in foo/urls.py
path("schema/", SpectacularAPIView.as_view(custom_settings=FOO_SPECTACULAR_SETTINGS), name="schema"),

# in bar/urls.py
path("schema/", SpectacularAPIView.as_view(custom_settings=BAR_SPECTACULAR_SETTINGS), name="schema"),

# in project_name/urls.py
path("foo/", include("foo.urls", namespace="foo")),
path("bar/", include("bar.urls", namespace="bar")),

Assuming that foo and bar have different enum overrides, loading foo first will break bar until the server restarts, and visa-versa.

Proposed fix
Make enum_name_overrides_setting an explicit argument to _load_enum_name_overrides, so it is cached properly:

def load_enum_name_overrides():
    return _load_enum_name_overrides(
      language=get_language(),
      enum_name_overrides_setting=spectacular_settings.ENUM_NAME_OVERRIDES,
    )

@functools.lru_cache()
def _load_enum_name_overrides(language: str, enum_name_overrides_setting: dict[str, Any]):
    overrides = {}
    for name, choices in enum_name_overrides_setting.items():
    ...
    # there is one more usage of the setting at the bottom of the function that also needs to be swapped

Note that using the full settings object itself as a cached function argument will NOT work, since its hash function is not impacted by the patching process.