Neil Conway <[EMAIL PROTECTED]> writes:
> On Fri, 2007-04-27 at 10:02 -0400, Tom Lane wrote:
>> Perhaps a sufficiently robust way would be to form the hash as the
>> XOR of each supplied digit, circular-shifted by say 3 times the
>> digit's weight.
> The only objection I have to this is that it means we need to have
> another hash function in the backend. The Jenkins hash we use in
> hash_any() has been studied and we can have at least some confidence in
> its collision-resistance, etc.
I'm still not very comfortable with that. You're proposing to add a
pretty obvious failure mechanism --- any numeric-returning function
that failed to "normalize" its output would now create a subtle,
hard-to-find bug. Even if you can promise that all the functions in
numeric.c get it right, what of user-written add-ons? And the only
return for taking this risk is speculation that the performance of the
hash function might be better.
I think if you want to go this way, you should at least provide some
evidence that we get a hashing performance benefit in exchange for
adding a new restriction on numeric-value validity. Perhaps a suitable
test would be to compare the number of hash collisions in a large set of
randomly-chosen-but-distinct numeric values.
regards, tom lane
---------------------------(end of broadcast)---------------------------
TIP 9: In versions below 8.0, the planner will ignore your desire to
choose an index scan if your joining column's datatypes do not