On Tue, Feb 27, 2018 at 2:08 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
Hadi Moshayedi <hadi@citusdata.com> writes: > I am wondering why is it not using index-only-scan (which would use the > cache better) and instead it does a bitmap scan?
Never experiment on an empty table and assume that the resulting plan is the same as you'd get on a large table.
In this case, not only don't you have any meaningful amount of data loaded, but the planner can see that none of the table's pages are marked all-visible, meaning that the "index-only" scan would degrade to a regular indexscan, which is how it gets costed. And on a single-page table, an indexscan is going to have a hard time beating other alternatives.
If one runs vacuum on a table (small or otherwise) that is currently choosing an index scan as its best plan how likely is it that post-vacuum an index-only plan would be chosen if the index type and column presence conditions are met?
Also, I recall discussion that select statements will touch the visibility map (hence causing write I/O even in a read-only query) but [1] indicates that only vacuum will set them ddl will clear them.