The 'Subscription List' SQL Problem

Question

Phil Factor SQL Speed Phreak Competition: No 1

This competition is now over, but the winner, Peso, got an Amazon Voucher for $60, and the privilege of being able to display the 'Phil Factor SQL Speed Phreak' award on their own site

It was quite a struggle with some close competition from many of those who participated in this competition. However, Peso came up with a blindingly fast winner that produced an aggregation from a million rows in a third of a second. Now, we're all scratching our heads trying to come up with the fastest way of solving....

The ‘FIFO Stock Inventory’ SQL Problem

...on this site

(here is the original preamble.)

I really genuinely don't know the answer to this question: What is the fastest way in SQL Server (any version) to provide the following subscription report.

It is a reasonable request. We have a subscription list with 10,000 subscribers and we need to do a management report that gives the monthly breakdown of the Date, the number of current subscribers at the end of the month, the number of resignations in the month ('Unsubscribes'), and the number of new subscribers. The list should be in date order, and the date should be just the date of the first day of the month.

The table is in this form (simplified from the way we'd do it in a real system of course)

CREATE TABLE [dbo].[Registrations]
    (
     [Registration_ID] [int] IDENTITY(1, 1)
                             NOT NULL,
     [FirstName] [varchar](80) NOT NULL,
     [LastName] [varchar](80) NOT NULL,
     [DateJoined] [datetime] NOT NULL,
     [DateLeft] [datetime] NULL,
     CONSTRAINT [PK_Registrations] PRIMARY KEY CLUSTERED 
    	([DateJoined], [LastName], [FirstName])
    )
CREATE INDEX idxDateJoined 
    ON Registrations (DateJoined, DateLeft, Registration_ID)

You are welcome to change the two indexes to suit your solution. I'll give you a reasonable amount of data to try stuff out on. 10,000 faked data entries of subscribers is the most that I can reasonably ask you to download, (The list is here) but I shall be taking each solution, putting it in in a test harness consisting of a million subscribers. (I may even make it 1,169,187 to celebrate SQLServerCentral's subscription list) , and find the fastest way of doing it. I have ideas of my own of the way to do this but I suspect they're wrong.

Note that in the sample data the subscriptions that extend to sept 2010 are those people who’ve paid for a year’s subscription only rather than those which have ongoing renewals (e.g. direct Debit). ‘Now’ is the end of September 2009.

I will allow you to use a number table, but you can assume that every month has new subscribers in it. You can use views, or temporary table, but the time taken for their creation will be included in the timings. You can use a cursor if you don't mind a sharp intake of breath from Jeff. You can use any version of SQL Server that you like.

The winner will be amongst the tied fastest entrants (generally there is a group of these) and it will be the one with the highest number of votes. We'll announce the winner in a week's time on 19th October.

Who knows, if you are the only person who ever discovers this site and this competition, then the winner will be you!

OK End of day 1 and we've had some extremely good results (in milliseconds) from the original 10,000 data table

 Matt 16 Kev Riley 93 Graham 60 Peso 33 Gianluca 16 AndyM 170 Joe Harris 15533 William brewer 406

I've had to temporarily disqualify Kev, AndyM and Joe Harris as I couldn't get their results to match the correct values. All the other results agreed. The results for Matt, Peso and Gianluca are all very close indeed, too close to get a result from 10,000 rows, so I'll have to up the test table to a million rows. From experience, I know that a lot could change between now and next Monday. I feel sure that Graham's SQL can be tweaked. and I expect to see the slow ones to be back in the running with some fine-tuning. Anyone else?

OK End of Day 2, and Peso has streaked ahead with his third version. it is now too fast to measure on 10000 rows so I shall have to move to a larger database for the performance tests. Now I've put result-checking into the test harness Gianluca has had to be temporarily disqualified until he can cure his result, which is slightly out.

Matt                 16 ms          
Kev Riley            76 ms *         
Graham               60 ms         
Peso                 30 ms         
Gianluca             33 ms *         
AndyM                170 ms *       
William brewer       406 ms        
Peso 2               16 ms         
Peso 3               0 ms

Result of SQL is incorrect
Day 3. Due to a couple of glitches I haven't been able to complete all the timings on the million-row table. Here is what I've done so far!

graham 1        6986ms  
Peso 1           890ms  
Peso 2          1173ms  
Peso 3          1170ms  
Matt 1           596ms  
Matt 2           873ms  
Gianluca 1      4546ms  
Peso 4B          940ms  
Andriy Z        1200ms *
graham 2        1656ms

Day 5

Currently re-creating the test harness! timings on the 1 Million row tables are as follows. remember that these are preliminary timings and Peso and I are trying to scratch our heads to see why we are getting such different results. In the test harness, the results are inserted into a table variable and subsequently checked for validity.

Entry       Elapsed time in milliseconds
graham 1    6983 ms
Peso 1       890 ms
Peso 2      1186 ms
Peso 3      1173 ms
Matt 1       576 ms
Matt 2       860 ms
Gianluca 1  4550 ms
Peso 4B      936 ms
Peso 4d      830 ms
Peso 4e      313 ms
Andriy Z    1203 ms
Graham 2    1640 ms
Brewer 2     406 ms
Peso 1d     1076 ms
Gustavo 1    580 ms
Gianluca 4  2390 ms

..at the moment I write this, William Brewer's entry (incorporating ideas from Graham, Peso and Matt) seems to be in the lead! (Saturday, it is now Peso 4e) However, I'll be holding the competition open until Monday evening GMT to allow for people who didn't hear that the competition was on last Monday.

Just if you thought that some potential winners were emerging, look at the results with the original 10,000 row table. Yes, quite different. (We'd have to use a C# or VB test-harness to sort out the speed differences!) One would never be able to sort out the fastest from such close competition!

Entry       Elapsed time in milliseconds
graham 1    80 ms
Peso 1      13
Peso 2      16
Peso 3      16
Matt 1      13
Matt 2      33
Gianluca 1  30
Peso 4B     16
Andriy Z    13
Graham 2    16
Brewer 2    30
Peso 1d     33
Gustavo 1   16
Gianluca 4  13
Peso 4d     16

The complete million long test table is now on http://www.simple-talk.com/blogbits/philf/registrations.zip if you would like to fine-tune your winning entry

Monday. Here are the final rankings These are the simple 'elapsed time' measurements. I'll follow this up as soon as possible with the details. Matt has very kindly written a special C# test harness for accurate timings that I'm dusting out at the moment. I'm planning to do a full write-up on my SSC blog as there are some important lessons for anyone doing this sort of reporting task. After all, there is a world of difference between the time and CPU loading of the best entrants and the Ho-hums. Even the Ho-hums are a lot better than some of the production code I've seen.

These timings are very tight so in all fairness, I have to also award Gustavo, as runner-up, the right to proudly display the 'Phil Factor SQL Speed Phreak' award. After all, what is 30 milliseconds in processing a million rows. Peso has to get the title, not only for the best entry (Peso 4E), but for his energy (he was on paternity leave!) and for the way he helped and advised the other entrants. A true champion. Special Thanks to Matt for all the advice, and the test harness.

Peso4E was such a good entry that it even struck terror into Celko-Prizewinner Barry, and decided him not to enter the competition. However William and Matt deserve special commendation for their entries which remain brilliantly fast over a million rows.

Peso 4E       313
Gustavo 3     343
Gustavo 4     346
Brewer 2      423
Peso 1e       470
Peso 1        500
Gustavo 1     563
Matt 1        580
Matt 1B       593
Peso 1d       856
GianLuca Final 856
Peso 4d       856
Peso 1D       860
Matt 2        873
Matt3         923
Peso 4B       940
Graham 2      1106
Peso 3        1156
Peso 2        1170
Andriy Z      1233 *
Gustavo/peso  2800
Gianluca 1    4500
graham 1      4656

Remember that the winning routines were calculating aggregate reports on a million-row tables in between a third and half a second. These included calculations that you will sometimes read has to be done using a cursor!

(* in results means result needs minor tweaking to make it conform.

Answer 1

I must admit that the key to the query is the UNPIVOT that Peso discovered to be the only way to perform a single scan instead of a double scan. Really impressive!

I know this will be disqualified, but I'm posting the quirky XML select anyway, modified with the UNPIVOT thing. Maybe it has a chance to perform fast on a 1M row table.

DECLARE @monthly_joins int
DECLARE @monthly_leaves int
DECLARE @running_users int
DECLARE @current_month date
DECLARE @xchar nvarchar(max)
DECLARE @xml XML
DECLARE @now date

SELECT @xchar = '', @monthly_joins = 0, @monthly_leaves = 0, @current_month = NULL, @running_users = 0, @now = '20091001'

SELECT 
    @xchar = 
    	CASE 
    		WHEN eventDate = ISNULL(@current_month, eventDate) 
    			THEN @xchar 
    		ELSE 
    			CASE
    				WHEN @current_month < @now 
    					THEN @xchar + '<A MONTH="'+ CONVERT(char(8),ISNULL(@current_month,eventDate),112) +'" JOINS="'+ CAST(@monthly_joins AS varchar(10)) +'" LEAVES="'+ CAST(@monthly_leaves AS varchar(10)) +'" TOT="'+ CAST(@running_users AS varchar(10)) +'"/>' 
    					ELSE @xchar
    				END
    	END,
    @monthly_joins = 
    	CASE 
    		WHEN eventDate = ISNULL(@current_month,eventDate) 
    			THEN @monthly_joins + JOINS 
    		ELSE JOINS 
    	END,
    @monthly_leaves = 
    	CASE 
    		WHEN eventDate = ISNULL(@current_month,eventDate) 
    			THEN @monthly_leaves + LEAVES
    		ELSE LEAVES
    	END,
    @running_users = @running_users + JOINS - LEAVES,
    @current_month = eventDate
FROM (
    	SELECT TOP 2147483647 CAST(CONVERT(char(6), eventDate, 112) + '01' AS DATE) AS eventDate, 
    		JOINS = CASE eventType WHEN 'DateJoined' THEN 1 ELSE 0 END, 
    		LEAVES = CASE eventType WHEN 'DateLeft' THEN 1 ELSE 0 END
    	FROM registrations AS reg
    	UNPIVOT (eventDate FOR eventType IN (DateJoined, DateLeft)) AS upvt
    	ORDER BY eventDate
) AS Data


IF @current_month < @now
    SELECT @xchar = @xchar + '<A MONTH="'+ CONVERT(char(8), @current_month, 112) +'" JOINS="'+ CAST(@monthly_joins AS varchar(10)) +'" LEAVES="'+ CAST(@monthly_leaves AS varchar(10)) +'" TOT="'+ CAST(@running_users AS varchar(10)) +'"/>'
SELECT @xml = CAST('<R>' + @xchar + '</R>' AS XML)


SELECT DISTINCT CONVERT(date, T.results.value('@MONTH[1]','char(6)') + '01', 112) as [month],
    T.results.value('@JOINS[1]','int') as joins,
    T.results.value('@LEAVES[1]','int') as leaves,
    T.results.value('@TOT[1]','int') as tot
FROM @xml.nodes('/R/A') AS T(results)

Answer 2

I wish I'd seen this sooner! I see today's the expiration date, but on the chance I've ducked in under the wire because of my time zone, here goes:

--------------------------------------------------------------
-- set up and populate numbers table; note it HAS A '0' ROW
CREATE TABLE #Numbers(nn smallint identity primary key);
SET IDENTITY_INSERT #Numbers ON
DECLARE @ii smallint; SET @ii=0;
WHILE @ii <=100 BEGIN
    INSERT INTO #Numbers (nn) VALUES(@ii)
    SET @ii = @ii + 1
END
SET IDENTITY_INSERT #Numbers OFF

--select * from #Numbers

--------------------------------------------------------------
-- join to the numbers table; 
-- we get joined only where DATEADD(month,NumberFromNumbers,'1/1/2004') is from 1/1/2004 thru LAST MONTH
-- NOTE this construct gets the first day of the month where SOMEDATE exists
--  	dateadd(month,datediff(month,0,SOMEDATE),0)
-- (This is a great example of a place where you WANT (for readability) to use a UDF but you SHOULDN'T (for performance))
--
SELECT 
    TheMonth,
    SUM(SubscriberCount) as SubscriberCount,
    SUM(ResigCount) as ResigCount,
    SUM(NewSubscriberCount) as NewSubscriberCount
FROM
    (
    SELECT 
    	DATEADD(month,nn.nn,'1/1/2004') as TheMonth,
    	-- Subscriber: DATEJOINED <= CURRENTMONTH AND (DATELEFT ISNULL or > CURRENTMONTH)
    	CASE WHEN dateadd(month,datediff(month,0,DateJoined),0) <= DATEADD(month,nn.nn,'1/1/2004') 
    			AND ISNULL(DateLeft,getdate()) > DATEADD(month,nn.nn,'1/1/2004') 
    		THEN 1 ELSE 0 END as SubscriberCount,
    	-- Resigned: DATELEFT = CURRENTMONTH
    	CASE WHEN dateadd(month,datediff(month,0,ISNULL(DateLeft,getdate())),0) = DATEADD(month,nn.nn,'1/1/2004') 
    		THEN 1 ELSE 0 END as ResigCount,
    	-- NewSubscriber: DATEJOINED = CURRENTMONTH
    	CASE WHEN dateadd(month,datediff(month,0,DateJoined),0) = DATEADD(month,nn.nn,'1/1/2004') 
    		THEN 1 ELSE 0 END as NewSubscriberCount	
    FROM registrations reg
    INNER JOIN #Numbers nn ON DATEADD(month,nn.nn,'1/1/2004') < dateadd(month,datediff(month,0,getdate()),0)
    ) xx
GROUP BY TheMonth
ORDER BY TheMonth

Answer 3

A second attempt, with a little more work put into it this time:

WITH MonthlySummaries([Date], joined, [left])--a table collected from two aggregations, giving us
  AS (SELECT    joined.[date],--those who joined and those who left in each month
                joined.joined,
                COALESCE([left].[left], 0)
      FROM      (SELECT DATEADD(MONTH, DATEDIFF(MONTH, 0, Registrations.[DateJoined]), 0)as [date],
                        COUNT(*) AS [joined]
                 FROM   Registrations
                 GROUP BY DATEADD(MONTH, DATEDIFF(MONTH, 0, Registrations.[DateJoined]), 0)
                ) joined
                LEFT OUTER JOIN (SELECT COUNT(*) AS [left],
                                        DATEADD(MONTH, DATEDIFF(MONTH, 0, Registrations.[Dateleft]), 0) as [date]
                                 FROM   Registrations
                                 GROUP BY DATEADD(MONTH, DATEDIFF(MONTH, 0, Registrations.[Dateleft]), 0)
                                ) [left] ON joined.date=[left].date
     )

SELECT a.date, min(a.joined) as [joined], min(a.[left]) as [left], SUM(b.joined)-SUM(b.[left]) AS subscribers FROM MonthlySummaries a INNER JOIN Monthlysummaries b ON b.date<=a.date GROUP BY a.date ORDER BY a.date

Answer 4

0

Heh... careful Peso... the Triangular Join only ran on the final 69 rows. It was all the other good stuff before the final running total that made it so fast. The Triangular Join would have beat the tar out of the server if you had run it against the original 10,000 rows. Imagine what it would have done on a million rows.

answered Feb 01 '10 at 10:37 AM

Jeff Moden
1k ● 1 ● 3

(comments are locked)

Answer 5

0

Phil... I love this. There are basically "no rules" except to meet the requirements in T-SQL. Contests like this are going to lead to a whole lot of innovation by a whole lot of folks. Thanks for doing this.

answered Feb 01 '10 at 10:39 AM