12

Given the following data:

create table #histories
(
    username varchar(10),
    account varchar(10),
    assigned date  
);

insert into #histories 
values 
('PHIL','ACCOUNT1','2017-01-04'),
('PETER','ACCOUNT1','2017-01-15'),
('DAVE','ACCOUNT1','2017-03-04'),
('ANDY','ACCOUNT1','2017-05-06'),
('DAVE','ACCOUNT1','2017-05-07'),
('FRED','ACCOUNT1','2017-05-08'),
('JAMES','ACCOUNT1','2017-08-05'),
('DAVE','ACCOUNT2','2017-01-02'),
('PHIL','ACCOUNT2','2017-01-18'),
('JOSH','ACCOUNT2','2017-04-08'),
('JAMES','ACCOUNT2','2017-04-09'),
('DAVE','ACCOUNT2','2017-05-06'),
('PHIL','ACCOUNT2','2017-05-07') ; 

... which represents when a given user was assigned to an account.

I am looking to establish who owned a given account on the last day of each month (the assigned date is the date that the account transferred ownership), with any missing month-ends populated (possibly created from a handy dates table that I have available, with useful columns DateKey, Date and LastDayOfMonth, [courtesy of @AaronBertrand])1.

The desired results would be:

PETER, ACCOUNT1, 2017-01-31
PETER, ACCOUNT1, 2017-02-28
DAVE, ACCOUNT1, 2017-03-31
DAVE, ACCOUNT1, 2017-04-30
FRED, ACCOUNT1, 2017-05-31
FRED, ACCOUNT1, 2017-06-30
FRED, ACCOUNT1, 2017-07-31
JAMES, ACCOUNT1, 2017-08-31
PHIL, ACCOUNT2, 2017-01-31
PHIL, ACCOUNT2, 2017-02-28
PHIL, ACCOUNT2, 2017-03-31
JAMES, ACCOUNT2, 2017-04-30
PHIL, ACCOUNT2, 2017-05-31

Doing the initial part of this with a windowing function is trivial, it's adding the "missing" rows that I'm struggling with.

  • So you're assuming that phil has the account on the last day of 2017-05 because he had it on 2017-05-07 and there was no subsequent holder? – Evan Carroll Aug 9 '17 at 16:22
  • Yes, that is the logic – Philᵀᴹ Aug 9 '17 at 16:24
9
+500

One approach to this problem is to do the following:

  1. Emulate LEAD on SQL Server 2008. You can use APPLY or a suquery for this.
  2. For rows without a next row, add one month to their account date.
  3. Join to a dimension table that contains month end dates. This eliminates all rows that don't span at least a month and adds rows to fill in the gaps as necessary.

I modified your test data a little bit to make the results deterministic. Also added an index:

create table #histories
(
    username varchar(10),
    account varchar(10),
    assigned date  
);

insert into #histories 
values 
('PHIL','ACCOUNT1','2017-01-04'),
('PETER','ACCOUNT1','2017-01-15'),
('DAVE','ACCOUNT1','2017-03-04'),
('ANDY','ACCOUNT1','2017-05-06'),
('DAVE','ACCOUNT1','2017-05-07'),
('FRED','ACCOUNT1','2017-05-08'),
('JAMES','ACCOUNT1','2017-08-05'),
('DAVE','ACCOUNT2','2017-01-02'),
('PHIL','ACCOUNT2','2017-01-18'),
('JOSH','ACCOUNT2','2017-04-08'), -- changed this date to have deterministic results
('JAMES','ACCOUNT2','2017-04-09'),
('DAVE','ACCOUNT2','2017-05-06'),
('PHIL','ACCOUNT2','2017-05-07') ;

-- make life easy
create index gotta_go_fast ON #histories (account, assigned);

Here's the laziest date dimension table of all time:

create table #date_dim_months_only (
    month_date date,
    primary key (month_date)
);

-- put 2500 month ends into table
INSERT INTO #date_dim_months_only WITH (TABLOCK)
SELECT DATEADD(DAY, -1, DATEADD(MONTH, ROW_NUMBER() OVER (ORDER BY (SELECT NULL)), '20000101'))
FROM master..spt_values;

For step 1, there are plenty of ways to emulate LEAD. Here's one method:

SELECT 
  h1.username
, h1.account
, h1.assigned
, next_date.assigned
FROM #histories h1
OUTER APPLY (
    SELECT TOP 1 h2.assigned
    FROM #histories h2
    WHERE h1.account = h2.account
    AND h1.assigned < h2.assigned
    ORDER BY h2.assigned ASC
) next_date;

For step 2, we need to change the NULL values to something else. You want to include the final month for each account, so adding one month to the starting date suffices:

ISNULL(next_date.assigned, DATEADD(MONTH, 1, h1.assigned))

For step 3, we can join to the date dimension table. The column from the dimension table is exactly the column you need for the result set:

INNER JOIN #date_dim_months_only dd ON
    dd.month_date >= h1.assigned AND
    dd.month_date < ISNULL(next_date.assigned, DATEADD(MONTH, 1, h1.assigned))

I didn't like the query that I got when I put it all together. There can be issues with join order when combining OUTER APPLY and INNER JOIN. To get the join order I wanted I rewrote it with a subquery:

SELECT 
  hist.username
, hist.account
, dd.month_date 
FROM
(
    SELECT 
      h1.username
    , h1.account
    , h1.assigned
    , ISNULL(
        (SELECT TOP 1 h2.assigned
            FROM #histories h2
            WHERE h1.account = h2.account
            AND h1.assigned < h2.assigned
            ORDER BY h2.assigned ASC
        )
        , DATEADD(MONTH, 1, h1.assigned)
    ) next_assigned
    FROM #histories h1
) hist
INNER JOIN #date_dim_months_only dd ON
    dd.month_date >= hist.assigned AND
    dd.month_date < hist.next_assigned;

I don't know how much data you have so it might not matter for you. But the plan looks how I want it to:

good plan

The results match yours:

╔══════════╦══════════╦════════════╗
║ username ║ account  ║ month_date ║
╠══════════╬══════════╬════════════╣
║ PETER    ║ ACCOUNT1 ║ 2017-01-31 ║
║ PETER    ║ ACCOUNT1 ║ 2017-02-28 ║
║ DAVE     ║ ACCOUNT1 ║ 2017-03-31 ║
║ DAVE     ║ ACCOUNT1 ║ 2017-04-30 ║
║ FRED     ║ ACCOUNT1 ║ 2017-05-31 ║
║ FRED     ║ ACCOUNT1 ║ 2017-06-30 ║
║ FRED     ║ ACCOUNT1 ║ 2017-07-31 ║
║ JAMES    ║ ACCOUNT1 ║ 2017-08-31 ║
║ PHIL     ║ ACCOUNT2 ║ 2017-01-31 ║
║ PHIL     ║ ACCOUNT2 ║ 2017-02-28 ║
║ PHIL     ║ ACCOUNT2 ║ 2017-03-31 ║
║ JAMES    ║ ACCOUNT2 ║ 2017-04-30 ║
║ PHIL     ║ ACCOUNT2 ║ 2017-05-31 ║
╚══════════╩══════════╩════════════╝
  • 500k rows. It's part of a nightly ETL, so it doesn't need to run in a millisecond :) – Philᵀᴹ Aug 10 '17 at 12:32
4

Here I don't use calendar table but a natural numbers table nums.dbo.nums (I hope you've got it too, if not, it can be easily generated)

I have the answer slightly different from yours ('JOSH' <-> 'JAMES') because your data contains these 2 rows:

('JOSH','ACCOUNT2','2017-04-09'),
('JAMES','ACCOUNT2','2017-04-09'),

with the same account and assigned date and you did not precise which one should be taken is this situation.

declare @eom table(account varchar(10), dt date); 

with acc_mm AS
(
select account, min(assigned) as min_dt, max(assigned) as max_dt
from #histories
group by account
),

acc_mm1 AS
(
select account,
       dateadd(month, datediff(month, '19991231', min_dt), '19991231') as start_dt,
       dateadd(month, datediff(month, '19991231', max_dt), '19991231') as end_dt
from acc_mm
)

insert into @eom (account, dt) 
select account, dateadd(month, n - 1, start_dt)
from acc_mm1
      join nums.dbo.nums            
           on n - 1 <= datediff(month, start_dt, end_dt); 

select eom.dt, eom.account, a.username
from @eom eom 
     cross apply(select top 1 *
                 from #histories h 
                 where eom.account = h.account
                   and h.assigned <= eom.dt
                 order by h.assigned desc) a
order by eom.account, eom.dt;                          
2

This is by no means a clean looking solution, but it seems to provide the results that you are looking for (I'm sure that others will have nice, clean, fully optimized queries for you).

create table #histories
(
    username varchar(10),
    account varchar(10),
    assigned date  
);

insert into #histories 
values 
('PHIL','ACCOUNT1','2017-01-04'),
('PETER','ACCOUNT1','2017-01-15'),
('DAVE','ACCOUNT1','2017-03-04'),
('ANDY','ACCOUNT1','2017-05-06'),
('DAVE','ACCOUNT1','2017-05-07'),
('FRED','ACCOUNT1','2017-05-08'),
('JAMES','ACCOUNT1','2017-08-05'),
('DAVE','ACCOUNT2','2017-01-02'),
('PHIL','ACCOUNT2','2017-01-18'),
('JOSH','ACCOUNT2','2017-04-09'),
('JAMES','ACCOUNT2','2017-04-09'),
('DAVE','ACCOUNT2','2017-05-06'),
('PHIL','ACCOUNT2','2017-05-07') ; 


IF (SELECT OBJECT_ID(N'tempdb..#IncompleteResults')) IS NOT NULL
    DROP TABLE #IncompleteResults;

DECLARE @EOMTable TABLE ( EndOfMonth DATE );
DECLARE @DateToWrite DATE = '2017-01-31';
WHILE @DateToWrite < '2017-10-31'
    BEGIN
        INSERT  INTO @EOMTable
                ( EndOfMonth )
                SELECT  @DateToWrite;

        SELECT  @DateToWrite = EOMONTH(DATEADD(MONTH, 1, @DateToWrite));
    END

    ;
WITH    cteAccountsByMonth
          AS ( SELECT   EndOfMonth ,
                        account
               FROM     @EOMTable e
                        CROSS JOIN ( SELECT DISTINCT
                                            account
                                     FROM   #histories
                                   ) AS h
             ),
        cteHistories
          AS ( SELECT   username ,
                        account ,
                        ROW_NUMBER() OVER ( PARTITION BY ( CAST(DATEPART(YEAR,
                                                              assigned) AS CHAR(4))
                                                           + ( RIGHT('00'
                                                              + CAST(DATEPART(MONTH,
                                                              assigned) AS VARCHAR(10)),
                                                              2) ) ), account ORDER BY assigned DESC ) AS rownum ,
                        CAST(DATEPART(YEAR, assigned) AS CHAR(4)) + RIGHT('00'
                                                              + CAST(DATEPART(MONTH,
                                                              assigned) AS VARCHAR(10)),
                                                              2) AS PartialDate ,
                        assigned ,
                        EOMONTH(assigned) AS EndofMonth
               FROM     #histories
             )
    SELECT  username ,
            e.EndOfMonth ,
            e.account
    INTO #IncompleteResults
    FROM    cteAccountsByMonth e
            LEFT JOIN cteHistories c ON e.EndOfMonth = c.EndofMonth
                                        AND c.account = e.account
                                        AND c.rownum = 1
SELECT  CASE WHEN username IS NULL
             THEN ( SELECT  username
                    FROM    #IncompleteResults i2
                    WHERE   username IS NOT NULL
                            AND i.account = i2.account
                            AND i2.EndOfMonth = ( SELECT    MAX(EndOfMonth)
                                                  FROM      #IncompleteResults i3
                                                  WHERE     i3.EndOfMonth < i.EndOfMonth
                                                            AND i3.account = i.account
                                                            AND i3.username IS NOT NULL
                                                )
                  )
             ELSE username
        END AS username ,
        EndOfMonth ,
        account 
FROM    #IncompleteResults i
ORDER BY account ,
        i.EndOfMonth;
2

I used the date dimension table from Aaron Bertrand, as you also mention in your question (which is a super-handy table for such scenarios) and I wrote the following code:

I added the EndOfMonth column to the #dim table(right after the FirstOfMonth column) using the following code:

 EndOfMonth as dateadd(s,-1,dateadd(mm, datediff(m,0,[date])+1,0)),

And the solution:

if object_id('tempdb..#temp') is not null drop table #temp
create table #temp (nr int, username varchar(100), account varchar(100), eom date)

;with lastassignedpermonth as
(
    select 
           month(assigned) month
         , account
         , max(assigned) assigned
    from 
           #histories 
    group by month(assigned), account 
)
insert into #temp
select 
       distinct row_number() over (order by d.account, d.eom) nr
     , h.username
     , d.account
     , d.eom
from ( 
        select distinct month, cast(d.endofmonth as date) eom, t.account 
        from #dim d cross apply (select distinct account from #histories) t 
     ) d
            left join lastassignedpermonth l on d.month = l.month and l.assigned <= d.eom and d.account = l.account 
            left join #histories h on h.assigned = l.assigned and h.account = l.account 
where d.eom <=  dateadd(s,-1,dateadd(mm, datediff(m,0,getdate())+1,0)) -- end of current month
order by d.account, eom 

-- This could have been done in one single statement with the lead() function but that is available as of SQL Server 2012
select case when t.username is null then (select username from #temp where nr = previous_username.nr) else t.username end as username, t.account, t.eom 
from #temp as t cross apply ( 
                                select max(nr) nr 
                                from #temp as t1
                                where t1.nr < t.nr and t1.username is not null
                            ) as previous_username

/*
   Note: You get twice JAMES and JOSH for April/ACCOUNT2, because apparently they are both assigned on the same date(2017-04-09)... 
   I guess your data should be cleaned up of overlapping dates.
*/
2

Triangle JOIN for the win!

SELECT account,EndOfMonth,username
FROM (
    SELECT Ends.*, h.*
        ,ROW_NUMBER() OVER (PARTITION BY h.account,Ends.EndOfMonth ORDER BY h.assigned DESC) AS RowNumber
    FROM (
        SELECT [Year],[Month],MAX(DATE) AS EndOfMonth
        FROM #dim
        GROUP BY [Year],[Month]
        ) Ends
    CROSS JOIN (
        SELECT account, MAX(assigned) AS MaxAssigned
        FROM #histories
        GROUP BY account
        ) ac
    JOIN #histories h ON h.account = ac.account
        AND Year(h.assigned) = ends.[Year]
        AND Month(h.assigned) <= ends.[Month] --triangle join for the win!
        AND EndOfMonth < DATEADD(month, 1, Maxassigned)
    ) Results
WHERE RowNumber = 1
ORDER BY account,EndOfMonth;

Results are:

account     EndOfMonth  username

ACCOUNT1    2017-01-31  PETER
ACCOUNT1    2017-02-28  PETER
ACCOUNT1    2017-03-31  DAVE
ACCOUNT1    2017-04-30  DAVE
ACCOUNT1    2017-05-31  FRED
ACCOUNT1    2017-06-30  FRED
ACCOUNT1    2017-07-31  FRED
ACCOUNT1    2017-08-31  JAMES

ACCOUNT2    2017-01-31  PHIL
ACCOUNT2    2017-02-28  PHIL
ACCOUNT2    2017-03-31  PHIL
ACCOUNT2    2017-04-30  JAMES
ACCOUNT2    2017-05-31  PHIL

Interactive Execution Plan here.

I/O and TIME stats (truncated all the zero-values after logical reads):

(13 row(s) affected)

Table 'Worktable'.  Scan count 3, logical reads 35.
Table 'Workfile'.   Scan count 0, logical reads  0.
Table '#dim'.       Scan count 1, logical reads  4.
Table '#histories'. Scan count 1, logical reads  1.

SQL Server Execution Times:
    CPU time = 0 ms,  elapsed time = 3 ms.

Query to create required 'temp tables and test the T-SQL statement I'm suggesting:

IF OBJECT_ID('tempdb..#histories') IS NOT NULL
    DROP TABLE #histories

CREATE TABLE #histories (
    username VARCHAR(10)
    ,account VARCHAR(10)
    ,assigned DATE
    );

INSERT INTO #histories
VALUES
('PHIL','ACCOUNT1','2017-01-04'),
('PETER','ACCOUNT1','2017-01-15'),
('DAVE','ACCOUNT1','2017-03-04'),
('ANDY','ACCOUNT1','2017-05-06'),
('DAVE','ACCOUNT1','2017-05-07'),
('FRED','ACCOUNT1','2017-05-08'),
('JAMES','ACCOUNT1','2017-08-05'),
('DAVE','ACCOUNT2','2017-01-02'),
('PHIL','ACCOUNT2','2017-01-18'),
('JOSH','ACCOUNT2','2017-04-08'),
('JAMES','ACCOUNT2','2017-04-09'),
('DAVE','ACCOUNT2','2017-05-06'),
('PHIL','ACCOUNT2','2017-05-07');

DECLARE @StartDate DATE = '20170101'
    ,@NumberOfYears INT = 2;

-- prevent set or regional settings from interfering with 
-- interpretation of dates / literals
SET DATEFIRST 7;
SET DATEFORMAT mdy;
SET LANGUAGE US_ENGLISH;

DECLARE @CutoffDate DATE = DATEADD(YEAR, @NumberOfYears, @StartDate);

-- this is just a holding table for intermediate calculations:
IF OBJECT_ID('tempdb..#dim') IS NOT NULL
    DROP TABLE #dim

CREATE TABLE #dim (
    [date] DATE PRIMARY KEY
    ,[day] AS DATEPART(DAY, [date])
    ,[month] AS DATEPART(MONTH, [date])
    ,FirstOfMonth AS CONVERT(DATE, DATEADD(MONTH, DATEDIFF(MONTH, 0, [date]), 0))
    ,[MonthName] AS DATENAME(MONTH, [date])
    ,[week] AS DATEPART(WEEK, [date])
    ,[ISOweek] AS DATEPART(ISO_WEEK, [date])
    ,[DayOfWeek] AS DATEPART(WEEKDAY, [date])
    ,[quarter] AS DATEPART(QUARTER, [date])
    ,[year] AS DATEPART(YEAR, [date])
    ,FirstOfYear AS CONVERT(DATE, DATEADD(YEAR, DATEDIFF(YEAR, 0, [date]), 0))
    ,Style112 AS CONVERT(CHAR(8), [date], 112)
    ,Style101 AS CONVERT(CHAR(10), [date], 101)
    );

-- use the catalog views to generate as many rows as we need

INSERT #dim ([date])
SELECT d
FROM (
    SELECT d = DATEADD(DAY, rn - 1, @StartDate)
    FROM (
        SELECT TOP (DATEDIFF(DAY, @StartDate, @CutoffDate)) rn = ROW_NUMBER() OVER (
                ORDER BY s1.[object_id]
                )
        FROM sys.all_objects AS s1
        CROSS JOIN sys.all_objects AS s2
        -- on my system this would support > 5 million days
        ORDER BY s1.[object_id]
        ) AS x
    ) AS y;

/* The actual SELECT statement to get the results we want! */

SET STATISTICS IO, TIME ON;

SELECT account,EndOfMonth,username
FROM (
    SELECT Ends.*, h.*
        ,ROW_NUMBER() OVER (PARTITION BY h.account,Ends.EndOfMonth ORDER BY h.assigned DESC) AS RowNumber
    FROM (
        SELECT [Year],[Month],MAX(DATE) AS EndOfMonth
        FROM #dim
        GROUP BY [Year],[Month]
        ) Ends
    CROSS JOIN (
        SELECT account, MAX(assigned) AS MaxAssigned
        FROM #histories
        GROUP BY account
        ) ac
    JOIN #histories h ON h.account = ac.account
        AND Year(h.assigned) = ends.[Year]
        AND Month(h.assigned) <= ends.[Month] --triangle join for the win!
        AND EndOfMonth < DATEADD(month, 1, Maxassigned)
    ) Results
WHERE RowNumber = 1
ORDER BY account,EndOfMonth;

SET STATISTICS IO, TIME OFF;

--IF OBJECT_ID('tempdb..#histories') IS NOT NULL DROP TABLE #histories
--IF OBJECT_ID('tempdb..#dim') IS NOT NULL DROP TABLE #dim

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.