changeset 1023:7a262ecae4f3

svnwrap: use get_file instead of list_props during replay This patch has an enormous effect on the time taken to import commits which add many files. Specifically, it makes them much faster for pulls over the network because it avoids creating a new connection to subversion for every added file. In my testing, it dropped the time taken to import a revision that adds ~6500 files from approximately 1 hour to 30 seconds. I believe this test is representative of real-world performance improvements.
author David Schleimer <dschleimer@fb.com>
date Tue, 18 Jun 2013 17:47:37 -0700
parents 1c9b1d0e0ba3
children 8feff33e387d
files hgsubversion/svnwrap/svn_swig_wrapper.py
diffstat 1 files changed, 8 insertions(+), 3 deletions(-) [+]
line wrap: on
line diff
--- a/hgsubversion/svnwrap/svn_swig_wrapper.py
+++ b/hgsubversion/svnwrap/svn_swig_wrapper.py
@@ -471,9 +471,14 @@ class SubversionRepo(object):
                 sf = f[l:]
                 if links[f] or execs[f]:
                     continue
-                props = self.list_props(sf, revision)
-                links[f] = props.get('svn:special') == '*'
-                execs[f] = props.get('svn:executable') == '*'
+                # The list_props API creates a new connection and then
+                # calls get_file for the remote file case.  It also
+                # creates a new connection to the subversion server
+                # every time it's called.  As a result, it's actually
+                # *cheaper* to call get_file than list_props here
+                data, mode = self.get_file(sf, revision)
+                links[f] = mode == 'l'
+                execs[f] = mode == 'x'
 
     def get_revision(self, revision, editor):
         ''' feed the contents of the given revision to the given editor '''