Re: Hash support for arrays

Поиск
Список
Период
Сортировка
От Dean Rasheed
Тема Re: Hash support for arrays
Дата
Msg-id AANLkTinF2=fbw0EKHEf+j0SCqutc8KWnF44V1C-=SVz7@mail.gmail.com
обсуждение исходный текст
Ответ на Re: Hash support for arrays  (Nicolas Barbier <nicolas.barbier@gmail.com>)
Ответы Re: Hash support for arrays  (Kenneth Marshall <ktm@rice.edu>)
Список pgsql-hackers
On 3 November 2010 09:24, Nicolas Barbier <nicolas.barbier@gmail.com> wrote:
> 2010/11/2 Kenneth Marshall <ktm@rice.edu>:
>
>> Given that our hash implimentation mixes the input data well (It does.
>> I tested it.) then a simple rotate-and-xor method is all that should
>> be needed to maintain all of the needed information. The original
>> hash function has done the heavy lifting in this case.
>
> Even with the perfect hash function for the elements, certain
> combinations of elements could still lead to massive collisions. E.g.,
> if repeated values are typical in the input data we are talking about,
> then the rotate-and-xor method would still lead to collisions between
> any array of the same values of certain lengths, regardless of the
> value. In Tom's implementation, as he mentioned before, those
> problematical lengths would be multiples of 32 (e.g., an array of 32
> 1s would collide with an array of 32 2s would collide with an array of
> 32 3s, etc).
>

Yeah, rotate-and-xor is a pretty weak hashing algorithm, since any
array of 32 identical elements will hash to either 0 or -1. Similarly
various permutations or multiples of that array length will cause it
to perform badly.

The multiply-by-m algorithm doesn't have that weakness, provided m is
chosen carefully. There are a couple of qualities a good algorithm
should possess:

1). The bits from the individual element hash values should be
distributed evenly so that no 2 different hash values would result in
the same contribution to the final value. This is easy to achieve -
just make sure that m is odd.

2). The way that each element's hash value bits are distributed should
be different from the way that every other element's hash value bits
are distributed. m=31 achieves this pretty well, although there are
plenty of other equally valid choices.

Regards,
Dean


В списке pgsql-hackers по дате отправления:

Предыдущее
От: Hannu Krosing
Дата:
Сообщение: Re: why does plperl cache functions using just a bool for is_trigger
Следующее
От: Peter Eisentraut
Дата:
Сообщение: Re: why does plperl cache functions using just a bool for is_trigger