Fixes the issue of handling NaN/None by using pandas Series in an optimized way.

django_pandas/utils.py

@@ -1,6 +1,5 @@
 # coding: utf-8
 
-from math import isnan
 from django.core.cache import cache
 from django.utils.encoding import force_text
 
@@ -31,19 +30,27 @@ def invalidate_signal_handler(sender, **kwargs):
 def replace_pk(model):
     base_cache_key = get_base_cache_key(model)
 
-    def inner(pk_list):
-        cache_keys = [None if isnan(pk) else base_cache_key % pk
-                      for pk in pk_list]
-        out_dict = cache.get_many(frozenset(cache_keys))
-        try:
-            return [None if k is None else out_dict[k] for k in cache_keys]
-        except KeyError:
-            out_dict = {
-                base_cache_key % obj.pk: force_text(obj)
-                for obj in model.objects.filter(pk__in={pk for pk in pk_list
-                                                        if not isnan(pk)})}
+    def get_cache_key_from_pk(pk):
+        return None if pk is None else base_cache_key % pk
+
+    def inner(pk_series):
+        pk_series = pk_series.where(pk_series.notnull(), None)
+        cache_keys = pk_series.apply(
+            get_cache_key_from_pk, convert_dtype=False)
+        unique_cache_keys = list(filter(None, cache_keys.unique()))
+
+        if not unique_cache_keys:
+            return pk_series
+
+        out_dict = cache.get_many(unique_cache_keys)
+
+        if len(out_dict) < len(unique_cache_keys):
+            out_dict = {base_cache_key % obj.pk: force_text(obj)
+                        for obj in model.objects.filter(
+                            pk__in=list(filter(None, pk_series.unique())))}
             cache.set_many(out_dict)
-            return list(map(out_dict.get, cache_keys))
+
+        return list(map(out_dict.get, cache_keys))
 
     return inner