Greg Stein wrote:
I thought the whole reason for variant results was to avoid O(N^2)
attacks against the hash table. You would be defeating that work.
No. They (APR) did not solved the problem.
For instance, if all keys have the same length,
the new seed approach will simply rotate
the buckets.
APR basically went down the obscure-and-pray-
really-hard road. So there is nothing to defeat
right now. An actual fix would e.g.
* use a hash implementation with O(1) space
and runtime guarantees, or
* use a bucket implementation that switches
from a linear list to a tree above 4 entries
The latter is the solution with the least impact
but APR does not have b-trees and such that
it could simply use here.
I see no reason to use a stable hash. If outputs are supposed to
stable, then the *presentation * layer should be stabilizing. Not our
core data structure.
Directory deltification making wordpress.org
go from 400+GB to 10GB *is* a reason.
Without stable hashes, we would need special
code for hash deltification.
I do agree that most parts of Subversion don't
*need* stable hashes. But having them makes
our life much easier by e.g. repeatable test runs.
This is the same kind of silliness that stsp made to our core data
structures. I still think that should be reverted in favor of
presentation changes.
If Canadian "internet" had supported sending
e-mail (not just receiving), I had posted this more
than a week ago and spared stsp this discussion.
Again, these are my reasons for using svn_hash__make:
* consistent behavior of SVN across different APR versions
* give devs time to check all the 500+ places that create
hashes throughout SVN for implicit assumptions on
ordering and such.
* performance improvement; particularly with directory-
or property-related operations
None of these points have been invalidated by
your "let's do it in post" suggestion.
Cheers,
-g
-- Stefan^2.
On May 19, 2012 8:46 PM, "Stefan Fuhrmann"
<stefanfuhrm...@alice-dsl.de <mailto:stefanfuhrm...@alice-dsl.de>> wrote:
Hi all,
svn_hash__make is basically apr_hash__make with
a hash function that will produce the same results
(element order etc) with APR 1.4.6 as well as earlier
APR versions. Without it, different test runs will result
in different results. As a bonus, we save some runtime.
I have a patch sitting on my machine that will do a
global replace of all 500+ locations where we create
APR hashes. Often, we need to add another #include.
There is no further change but the patch is large.
If there are no strong objections, I will commit the
change next weekend.
-- Stefan^2.
