On Wed, Aug 27, 2014 at 7:08 PM, Tatsuo Ishii <ishii@postgresql.org> wrote:
> While looking into a btree internal page using pg_filedump against an
> int4 index generated pgbench, I noticed that only item 2 has length 8,
> which indicates that the index tuple has only tuple header and has no
> index data. In my understanding this indicates that the item is used
> to represent a down link to a page. Question is, why the item is 2,
> not 1. I thought an index tuple indicating down link is always 1. Is
> this a sign that something goes wrong?
No. On a non-rightmost page, the "high key" item is physically first
(which is a bit odd, because it serves as a high-bound invariant on
the items that the page stores, but it's convenient to do it that way
for other reasons). On an internal page (that is also non-rightmost),
the second item (which is the first "real" item - i.e. the item which
P_FIRSTDATAKEY() returns) is just placeholder garbage. The reason for
that is noted above _bt_compare():
* CRUCIAL NOTE: on a non-leaf page, the first data key is assumed to be* "minus infinity": this routine will always
claimit is less than the* scankey. The actual key value stored (if any, which there probably isn't)* does not matter.
Thisconvention allows us to implement the Lehman and* Yao convention that the first down-link pointer is before the
firstkey.* See backend/access/nbtree/README for details.
--
Peter Geoghegan
