I've been skimming through the bitmap index patch...

A scan needs to access at least five pages:

1. B-tree index (root+others, depending on depth)
2. The auxiliary heap page
3. bitmap index meta page
4. LOV page
5. bitmap page

That seems like a lot of indirection. A high startup cost is probably ok for typical bitmap index use cases and most of the needed pages should stay in memory, but could we simplify this? Why do we need the auxiliary heap, couldn't we just store the blk+offset of the LOV item directly in the b-tree index item?

And instead of having separate LOV pages that store a number of LOV items, how about storing each LOV item on a page of it's own, and using the rest of the page to store the last chunk of the bitmap. That would eliminate one page access, but more importantly, maybe we could then get rid of all the bm_last_* attributes in BMLOVItemData that complicate the patch quite a bit, while preserving the performance.

  Heikki Linnakangas
  EnterpriseDB   http://www.enterprisedb.com

---------------------------(end of broadcast)---------------------------
TIP 5: don't forget to increase your free space map settings

Reply via email to