in reply to Re^8: Hash order randomization is coming, are you ready?
in thread Hash order randomization is coming, are you ready?

that I have assumed the "new" constraints as a matter of course ever since the randomisation fix for Algorithmic Complexity Attack that was (breifly???) implement in 5.8.1.

Alas not everyone has been as diligent as you. :-) It is surprising how many real bugs this found.

what has actually changed internally

Ok, first some history. In 5.8.1 a very similar patch the one I have been working on was implemented. It broke lots of stuff, which was considered unacceptable for a minor release. So a new implementation was done. This implementation actually supported two types of hash, and two seeds, one constant determined at build time, and one random per process. By default hashes would use the constant seed, but when Perl noticed too many collisions in a bucket it would trigger a "rehash" using a random per-process seed, which would cause the hash value of all of its keys to be reclaculatied and would as a byproduct cause the hash'es keys to be removed from the shared string table.

All of this consumed processing time, and added code complexity.

5.17.6 returned things to roughly where they were in 5.8.1. The rehash mechanism and all overheads associated with it are removed. The hash seed is randomly initialized per process. etc.

Somewhat related is the actual hash function in 5.17.6 is different from 5.17.5, and we probably will use a yet again different hash function in 5.18.

And if I have my way hashes will be randomized on a per hash level as well. (So every hash would have its own order, regardless of what keys it stores or the history of the hash.