Ask hash algortihm in partitioning table by hash

Question

hi all,
I wanted to ask about the hash algortihm in partitioning table by hash .How hash algorithm evenly distribute the data on each partisi??anybody know??

Answer 1

A simple calculation probably isn't going to be possible. The problem is that a simple hash function is unlikely to produce the nearly uniform data distribution and a function that produces the nearly uniform data distribution is unlikely to be particularly simple to compute. ORA_HASH has to balance the desire to be quick to compute against the desire to produce nearly uniform data. I don't know what algorithm Oracle picked or where relatively they made the trade off-- I expect that the algorithm may well change across releases.
Conceptually, though, if you had a perfect hash algorithm, it would provide a perfect "fingerprint" of the data and any valid input would have an equal a priori probability of being mapped anywhere in the valid output set-- just like a perfect encryption algorithm generates encrypted output that would appear totally random. If you can look at a piece of data and know that its hash was more likely to be in one part of the output set than another, that would mean that there would be an opportunity to attack the hash-- you could probabilistically reconstruct data if you knew the hash.
If you assume that ORA_HASH is a perfect hash (it probably isn't, but it's probably close enough for this analysis so long as the range is a power of 2), then for any given input, any output is equally likely. If you create a hash partitioned table with 8 partitions, that basically equates to asking ORA_HASH to generate a hash that is between 0 and 7 and puts the data in whichever partition comes out. Since each value is equally likely to be in any of the partitions, you'd expect that the data would be equally distributed. Unless of course there is a lot of repetition in the values in the column you are partitioning by-- the hash for two identical values is identical, so that sort of repetitive data would cause the data distribution to be unequal.
To take a simplistic example, let's assume you were hashing numbers and let's say you have 8 partitions. The simplest possible hash algorithm would be "mod 8". That is, if you insert a value x, insert it into the partiition (x mod 8). If x = 2, 2 mod 8 = 2, so put it in partition 2. If x = 10, 10 mod 8 = 2, so put it in partition 2. If x = 14, 14 mod 8 = 6 so put it in partition 6. If the numeric values that you insert are uniformly distributed (not unlikely if you're dealing with large amounts of data and very likely if you're dealing with sequence-generated primary keys), all 8 partitions will be roughly equally full.
Of course, this algorithm is far from perfect-- if all the data you are inserting is a power of 2, for example, then all the odd partitions would be empty and the even numbered partitions would be full. ORA_HASH needs to be quite a bit more complex than a simple mod N in order to provide more uniform mapping even if there are patterns in the input.
Justin