版本2：如何编写postgres sql查询，返回M +某个列的不同值，返回M +未知的总记录？

Question

This is a slightly harder version of what I posted earlier. I don't want to edit the original to ruin a perfectly good answer to the question I had at the time.

Suppose a table named tau.

tau:
A | B | C
----------    
2 - 1 - red
3 - 1 - rod
4 - 1 - rope
6 - 5 - red
7 - 5 - rap
8 - 5 - rod
9 - 10 -rod
11- 10 -road
12- 13 -rud

Column A is the primary key. It will be unique. Column B is a foreign key. In my database no integer key is ever the same across tables. Column C is not a key.

This table will have a lot of rows. There are many other columns, say column C for instance, that are indexed for search.

I want exactly M distinct values from column B. Lets say M = 2

Importantly, I am using column B in the expression (it's indexed too!) to determine which B's to return.

Normally, I would go

select distinct B From tau Where C like 'r_d' AND B < 13 Order By B Desc Limit 2

and I get

B
-----
10
5

This is the current state of affairs, now. But I want to upgrade to a new scenario:

When the expression on BC is satisfied and adds the record to the results pile (C like 'r_d' AND B < 13), I want to return column A as well as column B, while maintaining the restriction of 2 at most distinct column B values.

Another important condition is the solution must work for expressions in the where clause that return true for possibly different values of C.

Behold, the results I want.

A | B
------
9 - 10
8 - 5
6 - 5

The trouble is, the actual limit of records returned can be greater than M, or 2 in the example. I don't really care how many records come through, as long as there are only M distinct values of B. How would I write a query to accomplish this in Postgresql?

Answer 1

DENSE_RANK should do what you need. Here is SQL Fiddle . Put your M into WHERE R.rnk <= 2 instead of 2 .

Rather than starting with M distinct values from B and trying to figure out how to add missing values from column A to the final result, I'm thinking about this problem from the other side. We have some search conditions that limit the whole table to some intermediary result set:

SELECT *
FROM tau
WHERE C LIKE 'r_d' AND B < 13

This intermediary result set has all columns from the table, nothing is aggregated yet. We just need to filter this set further and leave only M distinct values of the B column. DENSE_RANK assigns a number (without gaps) to each group of B values, so we can use it in the final filter.

SELECT
    A, B
FROM
    (
        SELECT
            A
            ,B
            ,C
            ,DENSE_RANK() OVER(ORDER BY B DESC) AS rnk
        FROM tau
        WHERE C LIKE 'r_d' AND B < 13
    ) AS R
WHERE R.rnk <= 2
ORDER BY B DESC, A DESC;

result set

A    B
9    10
8    5
6    5

Second variant using LATERAL JOIN

SQL Fiddle with both variants. In the second variant we are finding M distinct values of B at first. Then for each found value we run the main query with extra filter for this B value. If there are millions of different values of B , then filtering to specific value should be efficient, provided there is an index on B . If there are millions of rows, but number of total distinct values of B is small, it should be slow.

WITH
CTE
AS
(
    select distinct B 
    From tau 
    Where C like 'r_d' AND B < 13 
    Order By B Desc 
    Limit 2
)
SELECT
  T.A, T.B, T.C
FROM
  CTE
  INNER JOIN LATERAL
  (
      SELECT tau.A, tau.B, tau.C
      FROM tau
      WHERE tau.B = CTE.B AND tau.C like 'r_d' AND tau.B < 13 
  ) AS T ON true
ORDER BY T.B DESC, T.A DESC;