On 16 October 2015 18:20:59 EEST, Bruce Momjian <bruce@momjian.us> wrote:
>I think on-disk bitmap indexes would only beat GIN indexes in a
>read-only database on low-cardinality columns. For example, if you had
>a purchase_log table and wanted to know all the "blue" and "large"
>items
>sold at a specific store, I can see on-disk bitmap indexes doing well
>there. If you added the product number, or needed read/write, I think
>GIN would win. I just don't think we have enough deployments who need
>what on-disk bitmap are best at.
My take on this is that we effectively already have bitmap indexes: it's called GIN. We could make the posting list
compressioneven better, currently a TID is compressed at best to a single byte, while in a bitmap index it could go
downto one bit, or even less. But that's just a matter of improving the compression algorithm, making it more bitmapy,
andcould be done as a fairly isolated change in GIN code.
Besides being more dense, there are some other tricks often associated with bitmap indexes. Instead of storing a
bitmap/postinglist per each unique value, you could store one for a range of values. That's useful e.g. for storing
floats,where you don't have many exact duplicate values, but you could get a dense index by treating all values in
range0.0-10.0 as one entry, all values in 10.0-50.0 as another, and so forth. Yet another trick is to have one bitmap
forall values > 0.0, another for all values >10.0, and so forth. With that, you can satisfy any BETWEEN query by
scanningjust two bitmaps/posting lists: the one for the lower bound and the one for the upper bound. The matching
tuplesare the ones that are present in the first, but not the latter posting list. But that's also not a whole new
indextype. GIN could do all that, if someone just wrote an opclass for it.
- Heikki
