--- a/docs/usage/performance.rst	Fri Aug 12 10:01:12 2016 +0200
+++ b/docs/usage/performance.rst	Tue Sep 06 00:51:18 2016 +0200
@@ -31,8 +31,26 @@
 
     Scaling horizontally can give huge performance benefits when dealing with
     large amounts of traffic (many users, CI servers, etc.). Kallithea can be
-    scaled horizontally on one (recommended) or multiple machines. In order
-    to scale horizontally you need to do the following:
+    scaled horizontally on one (recommended) or multiple machines.
+
+    It is generally possible to run WSGI applications multithreaded, so that
+    several HTTP requests are served from the same Python process at once. That
+    can in principle give better utilization of internal caches and less
+    process overhead.
+    
+    One danger of running multithreaded is that program execution becomes much
+    more complex; programs must be written to consider all combinations of
+    events and problems might depend on timing and be impossible to reproduce.
+
+    Kallithea can't promise to be thread-safe, just like the embedded Mercurial
+    backend doesn't make any strong promises when used as Kallithea uses it.
+    Instead, we recommend scaling by using multiple server processes.
+
+    Web servers with multiple worker processes (such as ``mod_wsgi`` with the
+    ``WSGIDaemonProcess`` ``processes`` parameter) will work out of the box.
+
+    In order to scale horizontally on multiple machines, you need to do the
+    following:
 
     - Each instance's ``data`` storage needs to be configured to be stored on a
       shared disk storage, preferably together with repositories. This ``data``
@@ -40,7 +58,7 @@
       task locking (so it is safe across multiple instances). Set the
       ``cache_dir``, ``index_dir``, ``beaker.cache.data_dir``, ``beaker.cache.lock_dir``
       variables in each .ini file to a shared location across Kallithea instances
-    - If celery is used each instance should run a separate Celery instance, but
+    - If using several Celery instances,
       the message broker should be common to all of them (e.g.,  one
       shared RabbitMQ server)
     - Load balance using round robin or IP hash, recommended is writing LB rules