[Python-Dev] Status of the fix for the hash collision vulnerability

Gregory P. Smith greg at krypto.org
Sat Jan 14 02:58:23 CET 2012

On Fri, Jan 13, 2012 at 5:38 PM, Guido van Rossum <guido at python.org> wrote:

> On Fri, Jan 13, 2012 at 5:17 PM, Antoine Pitrou <solipsis at pitrou.net>wrote:
>> On Thu, 12 Jan 2012 18:57:42 -0800
>> Guido van Rossum <guido at python.org> wrote:
>> > Hm... I started out as a big fan of the randomized hash, but thinking
>> more
>> > about it, I actually believe that the chances of some legitimate app
>> having
>> > >1000 collisions are way smaller than the chances that somebody's code
>> will
>> > break due to the variable hashing.
>> Breaking due to variable hashing is deterministic: you notice it as
>> soon as you upgrade (and then you use PYTHONHASHSEED to disable
>> variable hashing). That seems better than unpredictable breaking when
>> some legitimate collision chain happens.
> Fair enough. But I'm now uncomfortable with turning this on for bugfix
> releases. I'm fine with making this the default in 3.3, just not in 3.2,
> 3.1 or 2.x -- it will break too much code and organizations will have to
> roll back the release or do extensive testing before installing a bugfix
> release -- exactly what we *don't* want for those.
> FWIW, I don't believe in the SafeDict solution -- you never know which
> dicts you have to change.

Of the three options Victor listed only one is good.

I don't like *SafeDict*.  *-1*.  It puts the onerous on the coder to always
get everything right with regards to data that came from outside the
process never ending up hashed in a non-safe dict or set *anywhere*.
 "Safe" needs to be the default option for all hash tables.

I don't like the "*too many hash collisions*" exception. *-1*. It provides
non-deterministic application behavior for data driven applications with no
way for them to predict when it'll happen or where and prepare for it. It
may work in practice for many applications but is simply odd behavior.

I do like *randomly seeding the hash*. *+1*. This is easy. It can easily be
back ported to any Python version.

It is perfectly okay to break existing users who had anything depending on
ordering of internal hash tables. Their code was already broken. We
*will*provide a flag and/or environment variable that can be set to
turn the
feature off at their own peril which they can use in their test harnesses
that are stupid enough to use doctests with order dependencies.

This approach worked fine for Perl 9 years ago.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-dev/attachments/20120113/10c8ae68/attachment.html>

More information about the Python-Dev mailing list