On Fri, Jul 8, 2016 at 9:49 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Thu, Jul 7, 2016 at 4:56 PM, Bill Moran <wmoran@potentialtech.com> wrote:
>> SELECT DISTINCT size FROM grue;
>>
>> Always does a seq scan on Postgres 9.5.2. (Yes, I know we're
>> a patch behind, the upgrade is on the schedule) on
>> Ubuntu 14.
>>
>> I would expect it to be possible, and significantly more
>> efficient to do an index scan for that query.
>
> [...]
>
> We're probably missing a few tricks on queries of this type. If the
> index-traversal machinery had a mechanism to skip quickly to the next
> distinct value, that could be used here: walk up the btree until you
> find a page that contains keyspace not equal to the current key, then
> walk back down until you find the first leaf page that contains such a
> value. That would potentially let you step over large chunks of the
> index without actually examining all the leaf pages, which for a query
> like this seems like it could be a big win.
FWIW I messed around with prototyping this idea here:
https://www.postgresql.org/message-id/CADLWmXWALK8NPZqdnRQiPnrzAnic7NxYKynrkzO_vxYr8enWww@mail.gmail.com
I hope to return to that and some related ideas eventually as I learn
more about the relevant areas of the source code, if someone doesn't
beat me to it.
https://wiki.postgresql.org/wiki/Loose_indexscan shows a recursive CTE
that does the same thing at a higher level.
--
Thomas Munro
http://www.enterprisedb.com