Avoid conflict error hotspot in PluginIndexes' Unindex class by using...

Avoid conflict error hotspot in PluginIndexes' Unindex class by using IITreeSets instead of simple ints from the start. Idea taken from ``enfold.fixes``. This optimizes the common case of FieldIndexes being used for a low number of unique values. The edge-case of having a FieldIndex for an unique value per document gets penalized, as it uses many more persistent objects. This really warrants a separate type of index.

doc/CHANGES.rst

@@ -26,6 +26,10 @@ Bugs Fixed
 Features Added
 ++++++++++++++
 
+- Avoid conflict error hotspot in PluginIndexes' Unindex class by using
+  IITreeSets instead of simple ints from the start. Idea taken from
+  ``enfold.fixes``.
+
 - Added date range index improvements from ``experimental.catalogqueryplan``.
 
 - Changed policy on handling exceptions during ZCML parsing in ``Products``.

src/Products/PluginIndexes/common/UnIndex.py

@@ -204,22 +204,19 @@ class UnIndex(SimpleItem):
         indexRow = self._index.get(entry, _marker)
 
         # Make sure there's actually a row there already. If not, create
-        # an IntSet and stuff it in first.
+        # a set and stuff it in first.
         if indexRow is _marker:
-            self._index[entry] = documentId
-            # XXX _length needs to be migrated to Length object
-            try:
-                self._length.change(1)
-            except AttributeError:
-                if isinstance(self.__len__, Length):
-                    self._length = self.__len__
-                    del self.__len__
+            # We always use a set to avoid getting conflict errors on
+            # multiple threads adding a new row at the same time
+            self._index[entry] = IITreeSet((documentId, ))
             self._length.change(1)
         else:
-            try: indexRow.insert(documentId)
+            try:
+                indexRow.insert(documentId)
             except AttributeError:
-                # index row is an int
-                indexRow=IITreeSet((indexRow, documentId))
+                # Inline migration: index row with one element was an int at
+                # first (before Zope 2.13).
+                indexRow = IITreeSet((indexRow, documentId))
                 self._index[entry] = indexRow
 
     def index_object(self, documentId, obj, threshold=None):