I have a large table (200 million rows) with a column ( 'url' character varying(255)) that I need to be unique.
Currently I do this via a UNIQUE btree index on (lower(url::text)) The index is huge, and I would like to make it much smaller. Accesses to the table via this key are a tiny portion of the total (<1%), mainly being INSERTs which are not time critical. SELECTs very rarely use this column in WHERE, and never for time-critical queries. I would also like to partition the table, but this column and index is not a good choice for the partitioning criteria. I thought I might be able to create another column ('url_hash'), being a say 64-bit hash of lower(url::text) with a UNIQUE constraint. 64 bits should give me 36 bits of randomness over my 2^28 rows, making the probability of a false match sufficiently rare (probably much rarer than the risk of a programming bug causing a false match). Alternatively, I could use an even shorter hash (say 32 bits), and allow for the possibility of hash collisions. Does anybody know of any reference to using a hash in postgresql as a substitute for a unique btree? I would like to avoid re-inventing the wheel if possible. Stephen