Effective way of locating top ranked rows on Oracle DB - sql

Question

I have a large table (millions of records) and I need to write an efficient select statement.
The table looks like this:
create table tab1 (
pt_key number
, cp_key number
, ext_info varchar2(10)
, resp_nm varchar2(20)
, resp_dttm date
, rank number
);
Sample records:
insert into tab1 values (1,1,'info1','OK', to_date('01.03.18 17:00:00','DD.MM.RR HH24:MI'),1);
insert into tab1 values (1,1,'info2','FAILED', to_date('01.03.18 17:00:00','DD.MM.RR HH24:MI'),2);
insert into tab1 values (1,1,'info3','SENT', to_date('01.03.18 17:00:00','DD.MM.RR HH24:MI'),3);
insert into tab1 values (1,1,'info4','SENT', to_date('02.03.18 17:00:00','DD.MM.RR HH24:MI'),3);
insert into tab1 values (1,2,'info5','OK', to_date('05.03.18 17:00:00','DD.MM.RR HH24:MI'),1);
insert into tab1 values (1,2,'info6','OK', to_date('06.03.18 17:00:00','DD.MM.RR HH24:MI'),1);
insert into tab1 values (1,2,'info7','FAILED', to_date('01.03.18 17:00:00','DD.MM.RR HH24:MI'),2);
I would like the query to return for each combination of pt_key and cp_key (part of composite primary key, other columns are not indexed) record with the highest rank. If there are (for each combination of pt_key and cp_key) several records with the same highest rank then pick the one with the greatest resp_dttm.
The select statement should return only the first four columns.
For the above posted sample data the desired result would be:
1 1 info4 SENT
1 2 info7 FAILED
Thanks for help.

Answer 1

Here's one approach using row_number():
select *
from (
select *, row_number() over (partition by pt_key, cp_key
order by rank desc, resp_dttm desc) rn
from tab1
) t
where rn = 1

Answer 2

Here's another approach using FIRST aggregate function:
select pt_key,
cp_key,
max(ext_info) keep (dense_rank first order by t.rank desc, t.resp_dttm desc) as ext_info,
max(resp_nm) keep (dense_rank first order by t.rank desc, t.resp_dttm desc) as resp_nm
from tab1 t
group by pt_key, cp_key
Here's how it works on Oracle Live SQL
EDIT 2:
Result:
PT_KEY | CP_KEY | EXT_INFO | RESP_NM
--------+--------+----------+---------
1 | 1 | info4 | SENT
1 | 2 | info7 | FAILED
EDIT 1:
This solution has an important drawback, if for a certain combination of pt_key and cp_key, there are multiple rows with the same rank and resp_dttm values. In that case it will "combine" those rows, and calculate the aggregates for ext_info and resp_nm (in my example it'll take max value).
You can refine that behavior, by adding tertiary sort criteria, to make the ranking distinct (e.g. add all other columns from the primary key).
The answer from #sgeddes is a bit better in that sense, that it will use one (random) row from the equally ranked rows, without combining the data, and without having to add sorting criteria. It also is easier to maintain/update, as it has the ranking criteria in one place, while mine has it in two spots.
You should probably test performance of both in your specific scenario (e.g. specific indices, specific data profile/statistics).