Fixed a bug where South would cause Haystack to setup too soon. Thanks to adamfast for the report!

Author: toastdriven

haystack/models.py

@@ -1,5 +1,6 @@
 # "Hey, Django! Look at me, I'm an app! For Serious!"
 import logging
+from django.conf import settings
 from django.core.exceptions import ObjectDoesNotExist
 from django.db import models
 from django.utils.encoding import force_unicode
@@ -221,11 +222,46 @@ def __setstate__(self, data_dict):
 
 # Setup pre_save/pre_delete signals to make sure things like the signals in
 # ``RealTimeSearchIndex`` are setup in time to handle data changes.
-def load_indexes(sender, instance, *args, **kwargs):
+def load_indexes(sender, *args, **kwargs):
     from haystack import connections
 
     for conn in connections.all():
-        conn.get_unified_index().setup_indexes()
+        ui = conn.get_unified_index()
+        ui.setup_indexes()
+
+
+def reload_indexes(sender, *args, **kwargs):
+    from haystack import connections
+
+    for conn in connections.all():
+        ui = conn.get_unified_index()
+        # Note: Unlike above, we're resetting the ``UnifiedIndex`` here.
+        # Thi gives us a clean slate.
+        ui.reset()
+        ui.setup_indexes()
+
 
 models.signals.pre_save.connect(load_indexes, dispatch_uid='setup_index_signals')
 models.signals.pre_delete.connect(load_indexes, dispatch_uid='setup_index_signals')
+
+
+if 'south' in settings.INSTALLED_APPS:
+    # South causes a little mayhem, as when you run a ``syncdb``, it'll setup
+    # the apps *without* migrations using Django's built-in ``syncdb``. When
+    # this happens, ``INSTALLED_APPS`` consists of only those apps, NOT all
+    # apps.At the end of that sync, Django runs ``create_permissions``, which
+    # of course uses the ORM, causing the ``pre_save`` above to fire.
+
+    # The effect is that Haystack runs its setup against the then-subset of
+    # ``INSTALLED_APPS``. Once that's done, it won't re-setup the
+    # ``UnifiedIndex`` again, since the signal has a ``dispatch_uid``.
+
+    # This bug gets exposed only either when people run tests that *use*
+    # the South migrations OR when they have a data migration & the changed
+    # data isn't picked up by ``RealTimeSearchIndex`` (or similar).
+
+    # In the event of this, the only safe route is to listen for
+    # ``south.signals.post_migrate``, then re-run setup. This will
+    # unfortunately happen per-app, but should be quick & reliable.
+    from south.signals import post_migrate
+    post_migrate.connect(reload_indexes)