On Sat, Mar 03, 2018 at 06:01:23PM +0100, Branko Čibej wrote: > On 03.03.2018 17:44, Stefan Sperling wrote: > > On Sat, Mar 03, 2018 at 04:32:35PM +0000, Philip Martin wrote: > >> Stefan Sperling <s...@elego.de> writes: > >> > >>> Which leads me to believe that r1778923 may have been based on wrong > >>> assumptions about performance. The new authz is not fast enough to > >>> significantly reduce per-request overhead. > >> My testing so far was with a very small authz file -- only a handful of > >> rules and aliases. If I add a few hundred trival rules to the file then > >> 1.11 becomes signifcantly slower than 1.9 while reverting is still much > >> faster: > >> > >> 1.9: 4.3s > >> trunk 1.11: 14.6s > >> reverted 1.11: 1.9s > > Thanks for testing and confirming this. > > > > I think our best course of action is to revert the change on trunk > > and in 1.10.x. Could you do that? (I could do it, too. I'm just asking > > you since you've probably already prepared it in a local copy.) > > So if I understand this debate correctly: The authz code is so much > faster now that parsing the authz file and performing the authz lookups > beats calculating its MD5 checksum?
No, the file is only parsed/checksummed once. What's making it faster than 1.9 is that rule lookups happen on a cached copy of a parsed representation which is more efficient than the representation used in 1.9 (which uses svn_config_enumerate_sections2() to walk the ruleset). Re-reading the file and calculating its checksum is still slow and doing it per request really hurts if the file has a few hundred rules.
