I didn't notice Bloom filters in the conversation so far, and have been waiting for *years* for a good excuse to use a Bloom filter. I ran into them years back in Splunk, which is a distributed log store. There's an obvious benefit to a probabalistic tool like a Bloom filter there since remote lookup (and/or retrieval from cold storage) is quite expensive, relative to a local, hashed lookup. I haven't tried them in Postgres.
In the case of a single column with a small set of distinct values over a large set of rows, how would a Bloom filter be preferable to, say, a GIN index on an integer value? I have to say, this is actually a good reminder in my case. We've got a lot of small-distinct-values-big-rows columns. For example, "server_id", "company_id", "facility_id", and so on. Only a handful of parent keys with many millions of related rows. Perhaps it would be conceivable to use a Bloom index to do quick lookups on combinations of such values within the same table. I haven't tried Bloom indexes in Postgres, this might be worth some experimenting. Is there any thought in the Postgres world of adding something like Oracle's bitmap indexes?
