Query recursiva em SQL Server – Recursive CTE

For those not using SQL Server 2012, a cursor is likely the most efficient supported and guaranteedmethod outside of CLR. There are other approaches such as the “quirky update” which can be marginally faster but not guaranteed to work in the future, and of course set-based approaches with hyperbolic performance profiles as the table gets larger, and recursive CTE methods that often require direct #tempdb I/O or result in spills that yield roughly the same impact.


INNER JOIN – do not do this:

The slow, set-based approach is of the form:

SELECT t1.TID, t1.amt, RunningTotal = SUM(t2.amt) FROM dbo.Transactions AS t1 INNER JOIN dbo.Transactions AS t2 ON t1.TID >= t2.TID GROUP BY t1.TID, t1.amt ORDER BY t1.TID;

The reason this is slow? As the table gets larger, each incremental row requires reading n-1 rows in the table. This is exponential and bound for failures, timeouts, or just angry users.


Correlated subquery – do not do this either:

The subquery form is similarly painful for similarly painful reasons.

SELECT TID, amt, RunningTotal = amt + COALESCE( ( SELECT SUM(amt) FROM dbo.Transactions AS i WHERE i.TID < o.TID), 0 ) FROM dbo.Transactions AS o ORDER BY TID;

Quirky update – do this at your own risk:

The “quirky update” method is more efficient than the above, but the behavior is not documented, there are no guarantees about order, and the behavior might work today but could break in the future. I’m including this because it is a popular method and it is efficient, but that doesn’t mean I endorse it. The primary reason I even answered this question instead of closing it as a duplicate is because the other question has a quirky update as the accepted answer.

DECLARE @t TABLE ( TID INT PRIMARY KEY, amt INT, RunningTotal INT ); DECLARE @RunningTotal INT = 0; INSERT @t(TID, amt, RunningTotal) SELECT TID, amt, RunningTotal = 0 FROM dbo.Transactions ORDER BY TID; UPDATE @t SET @RunningTotal = RunningTotal = @RunningTotal + amt FROM @t; SELECT TID, amt, RunningTotal FROM @t ORDER BY TID;

Recursive CTEs

This first one relies on TID to be contiguous, no gaps:

;WITH x AS ( SELECT TID, amt, RunningTotal = amt FROM dbo.Transactions WHERE TID = 1 UNION ALL SELECT y.TID, y.amt, x.RunningTotal + y.amt FROM x INNER JOIN dbo.Transactions AS y ON y.TID = x.TID + 1 ) SELECT TID, amt, RunningTotal FROM x ORDER BY TID OPTION (MAXRECURSION 10000);

If you can’t rely on this, then you can use this variation, which simply builds a contiguous sequence using ROW_NUMBER():

;WITH y AS ( SELECT TID, amt, rn = ROW_NUMBER() OVER (ORDER BY TID) FROM dbo.Transactions ), x AS ( SELECT TID, rn, amt, rt = amt FROM y WHERE rn = 1 UNION ALL SELECT y.TID, y.rn, y.amt, x.rt + y.amt FROM x INNER JOIN y ON y.rn = x.rn + 1 ) SELECT TID, amt, RunningTotal = rt FROM x ORDER BY x.rn OPTION (MAXRECURSION 10000);

Depending on the size of the data (e.g. columns we don’t know about), you may find better overall performance by stuffing the relevant columns only in a #temp table first, and processing against that instead of the base table:

CREATE TABLE #x ( rn INT PRIMARY KEY, TID INT, amt INT ); INSERT INTO #x (rn, TID, amt) SELECT ROW_NUMBER() OVER (ORDER BY TID), TID, amt FROM dbo.Transactions; ;WITH x AS ( SELECT TID, rn, amt, rt = amt FROM #x WHERE rn = 1 UNION ALL SELECT y.TID, y.rn, y.amt, x.rt + y.amt FROM x INNER JOIN #x AS y ON y.rn = x.rn + 1 ) SELECT TID, amt, RunningTotal = rt FROM x ORDER BY TID OPTION (MAXRECURSION 10000); DROP TABLE #x;

Only the first CTE method will provide performance rivaling the quirky update, but it makes a big assumption about the nature of the data (no gaps). The other two methods will fall back and in those cases you may as well use a cursor (if you can’t use CLR and you’re not yet on SQL Server 2012).


Cursor

Everybody is told that cursors are evil, and that they should be avoided at all costs, but this actually beats the performance of most other supported methods, and is safer than the quirky update. The only ones I prefer over the cursor solution are the 2012 and CLR methods (below):

CREATE TABLE #x ( TID INT PRIMARY KEY, amt INT, rt INT ); INSERT #x(TID, amt) SELECT TID, amt FROM dbo.Transactions ORDER BY TID; DECLARE @rt INT, @tid INT, @amt INT; SET @rt = 0; DECLARE c CURSOR LOCAL STATIC READ_ONLY FORWARD_ONLY FOR SELECT TID, amt FROM #x ORDER BY TID; OPEN c; FETCH c INTO @tid, @amt; WHILE @@FETCH_STATUS = 0 BEGIN SET @rt = @rt + @amt; UPDATE #x SET rt = @rt WHERE TID = @tid; FETCH c INTO @tid, @amt; END CLOSE c; DEALLOCATE c; SELECT TID, amt, RunningTotal = rt FROM #x ORDER BY TID; DROP TABLE #x;

SQL Server 2012

In 2012, new window functions make this task a lot easier (and it performs better than all of the above methods as well):

SELECT TID, amt, RunningTotal = SUM(amt) OVER (ORDER BY TID ROWS UNBOUNDED PRECEDING) FROM dbo.Transactions ORDER BY TID;

Note that on larger data sets, you’ll find that the above performs much better than either of the following two options, since RANGE uses an on-disk spool (and the default uses RANGE). However it is also important to note that the behavior and results can differ, so be sure they both return correct results before deciding between them based on this difference.

SELECT TID, amt, RunningTotal = SUM(amt) OVER (ORDER BY TID) FROM dbo.Transactions ORDER BY TID; SELECT TID, amt, RunningTotal = SUM(amt) OVER (ORDER BY TID RANGE UNBOUNDED PRECEDING) FROM dbo.Transactions ORDER BY TID;

CLR

For completeness, I’m offering a link to Pavel Pawlowski’s CLR method, which is by far the preferable method on versions prior to SQL Server 2012 (but not 2000 obviously).

http://www.pawlowski.cz/2010/09/sql-server-and-fastest-running-totals-using-clr/


Conclusion

If you are on SQL Server 2012, the choice is obvious – use the new SUM() OVER() construct (with ROWS vs. RANGE). For earlier versions, you’ll want to compare the performance of the alternative approaches on your schema, data and – taking non-performance-related factors in mind – determine which approach is right for you. It very well may be the CLR approach. Here are my recommendations, in order of preference:

  1. SUM() OVER() ... ROWS, if on 2012
  2. CLR method, if possible
  3. First recursive CTE method, if possible
  4. Cursor
  5. The other recursive CTE methods
  6. Quirky update
  7. Join and/or correlated subquery

For further information with performance comparisons of these methods, see this question onhttp://dba.stackexchange.com:

http://dba.stackexchange.com/questions/19507/running-total-with-count