The most elegant way to generate permutations in SQL server

Question

Given a the following table:

Index | Element
---------------
  1   |    A
  2   |    B
  3   |    C
  4   |    D

We want to generate all the possible permutations (without repetitions) using the elements. the final result (skipping some rows) will look like this:

  Results
----------
   ABCD
   ABDC
   ACBD
   ACDB
   ADAC
   ADCA

   ...

   DABC
   DACB
   DBCA
   DBAC
   DCAB
   DCBA

  (24 Rows)

How would you do it?

Answer 1

After making some perhaps snarky comments, this problem stuck in my brain all evening, and I eventually came up with the following set-based approach. I believe it definitely qualifies as "elegant", but then I also think it qualifies as "kinda dumb". You make the call.

First, set up some tables:

--  For testing purposes
DROP TABLE Source
DROP TABLE Numbers
DROP TABLE Results


--  Add as many rows as need be processed--though note that you get N! (number of rows, factorial) results,
--  and that gets big fast. The Identity column must start at 1, or the algorithm will have to be adjusted.
--  Element could be more than char(1), though the algorithm would have to be adjusted again, and each element
--  must be the same length.
CREATE TABLE Source
 (
   SourceId  int      not null  identity(1,1)
  ,Element   char(1)  not null
 )

INSERT Source (Element) values ('A')
INSERT Source (Element) values ('B')
INSERT Source (Element) values ('C')
INSERT Source (Element) values ('D')
--INSERT Source (Element) values ('E')
--INSERT Source (Element) values ('F')


--  This is a standard Tally table (or "table of numbers")
--  It only needs to be as long as there are elements in table Source
CREATE TABLE Numbers (Number int not null)
INSERT Numbers (Number) values (1)
INSERT Numbers (Number) values (2)
INSERT Numbers (Number) values (3)
INSERT Numbers (Number) values (4)
INSERT Numbers (Number) values (5)
INSERT Numbers (Number) values (6)
INSERT Numbers (Number) values (7)
INSERT Numbers (Number) values (8)
INSERT Numbers (Number) values (9)
INSERT Numbers (Number) values (10)


--  Results are iteratively built here. This could be a temp table. An index on "Length" might make runs
--  faster for large sets.  Combo must be at least as long as there are characters to be permuted.
CREATE TABLE Results
 (
   Combo   varchar(10)  not null
  ,Length  int          not null
 )

Here's the routine:

SET NOCOUNT on

DECLARE
  @Loop     int
 ,@MaxLoop  int


--  How many elements there are to process
SELECT @MaxLoop = max(SourceId)
 from Source


--  Initialize first value
TRUNCATE TABLE Results
INSERT Results (Combo, Length)
 select Element, 1
  from Source
  where SourceId = 1

SET @Loop = 2

--  Iterate to add each element after the first
WHILE @Loop <= @MaxLoop
 BEGIN

    --  See comments below. Note that the "distinct" remove duplicates, if a given value
    --  is to be included more than once
    INSERT Results (Combo, Length)
     select distinct
        left(re.Combo, @Loop - nm.Number)
        + so.Element
        + right(re.Combo, nm.Number - 1)
       ,@Loop
      from Results re
       inner join Numbers nm
        on nm.Number <= @Loop
       inner join Source so
        on so.SourceId = @Loop
      where re.Length = @Loop - 1

    --  For performance, add this in if sets will be large
    --DELETE Results
    -- where Length <> @Loop

    SET @Loop = @Loop + 1
 END

--  Show results
SELECT *
 from Results
 where Length = @MaxLoop
 order by Combo

The general idea is: when adding a new element (say "B") to any string (say, "A"), to catch all permutations you would add B to all possible positions (Ba, aB), resulting in a new set of strings. Then iterate: Add a new element (C) to each position in a string (AB becomes Cab, aCb, abC), for all strings (Cba, bCa, baC), and you have the set of permutations. Iterate over each result set with the next character until you run out of characters... or resources. 10 elements is 3.6 million permutations, roughly 48MB with the above algorithm, and 14 (unique) elements would hit 87 billion permutations and 1.163 terabytes.

I'm sure it could eventually be wedged into a CTE, but in the end all that would be is a glorified loop. The logic is clearer this way, and I can't help but think the CTE execution plan would be a nightmare.