SQL to determine minimum sequential days of access?

Question

The following User History table contains one record for every day a given user has accessed a website (in a 24 hour UTC period). It has many thousands of r

Answer 1

This should do what you want but I don't have enough data to test efficiency. The convoluted CONVERT/FLOOR stuff is to strip the time portion off the datetime field. If you're using SQL Server 2008 then you could use CAST(x.CreationDate AS DATE).

DECLARE @Range as INT
SET @Range = 10

SELECT DISTINCT UserId, CONVERT(DATETIME, FLOOR(CONVERT(FLOAT, a.CreationDate)))
  FROM tblUserLogin a
WHERE EXISTS
   (SELECT 1 
      FROM tblUserLogin b 
     WHERE a.userId = b.userId 
       AND (SELECT COUNT(DISTINCT(CONVERT(DATETIME, FLOOR(CONVERT(FLOAT, CreationDate))))) 
              FROM tblUserLogin c 
             WHERE c.userid = b.userid 
               AND CONVERT(DATETIME, FLOOR(CONVERT(FLOAT, c.CreationDate))) BETWEEN CONVERT(DATETIME, FLOOR(CONVERT(FLOAT, a.CreationDate))) and CONVERT(DATETIME, FLOOR(CONVERT(FLOAT, a.CreationDate)))+@Range-1) = @Range)

Creation script

CREATE TABLE [dbo].[tblUserLogin](
    [Id] [int] IDENTITY(1,1) NOT NULL,
    [UserId] [int] NULL,
    [CreationDate] [datetime] NULL
) ON [PRIMARY]

Answer 2

The answer is obviously:

SELECT DISTINCT UserId
FROM UserHistory uh1
WHERE (
       SELECT COUNT(*) 
       FROM UserHistory uh2 
       WHERE uh2.CreationDate 
       BETWEEN uh1.CreationDate AND DATEADD(d, @days, uh1.CreationDate)
      ) = @days OR UserId = 52551

EDIT:

Okay here's my serious answer:

DECLARE @days int
DECLARE @seconds bigint
SET @days = 30
SET @seconds = (@days * 24 * 60 * 60) - 1
SELECT DISTINCT UserId
FROM (
    SELECT uh1.UserId, Count(uh1.Id) as Conseq
    FROM UserHistory uh1
    INNER JOIN UserHistory uh2 ON uh2.CreationDate 
        BETWEEN uh1.CreationDate AND 
            DATEADD(s, @seconds, DATEADD(dd, DATEDIFF(dd, 0, uh1.CreationDate), 0))
        AND uh1.UserId = uh2.UserId
    GROUP BY uh1.Id, uh1.UserId
    ) as Tbl
WHERE Conseq >= @days

EDIT:

[Jeff Atwood] This is a great fast solution and deserves to be accepted, but Rob Farley's solution is also excellent and arguably even faster (!). Please check it out too!

Answer 3

How about (and please make sure the previous statement ended with a semi-colon):

WITH numberedrows
     AS (SELECT ROW_NUMBER() OVER (PARTITION BY UserID 
                                       ORDER BY CreationDate)
                - DATEDIFF(day,'19000101',CreationDate) AS TheOffset,
                CreationDate,
                UserID
         FROM   tablename)
SELECT MIN(CreationDate),
       MAX(CreationDate),
       COUNT(*) AS NumConsecutiveDays,
       UserID
FROM   numberedrows
GROUP  BY UserID,
          TheOffset  

The idea being that if we have list of the days (as a number), and a row_number, then missed days make the offset between these two lists slightly bigger. So we're looking for a range that has a consistent offset.

You could use "ORDER BY NumConsecutiveDays DESC" at the end of this, or say "HAVING count(*) > 14" for a threshold...

I haven't tested this though - just writing it off the top of my head. Hopefully works in SQL2005 and on.

...and would be very much helped by an index on tablename(UserID, CreationDate)

Edited: Turns out Offset is a reserved word, so I used TheOffset instead.

Edited: The suggestion to use COUNT(*) is very valid - I should've done that in the first place but wasn't really thinking. Previously it was using datediff(day, min(CreationDate), max(CreationDate)) instead.

Rob

Answer 4

assuming a schema that goes like:

create table dba.visits
(
    id  integer not null,
    user_id integer not null,
    creation_date date not null
);

this will extract contiguous ranges from a date sequence with gaps.

select l.creation_date  as start_d, -- Get first date in contiguous range
    (
        select min(a.creation_date ) as creation_date 
        from "DBA"."visits" a 
            left outer join "DBA"."visits" b on 
                   a.creation_date = dateadd(day, -1, b.creation_date ) and 
                   a.user_id  = b.user_id 
            where b.creation_date  is null and
                  a.creation_date  >= l.creation_date  and
                  a.user_id  = l.user_id 
    ) as end_d -- Get last date in contiguous range
from  "DBA"."visits" l
    left outer join "DBA"."visits" r on 
        r.creation_date  = dateadd(day, -1, l.creation_date ) and 
        r.user_id  = l.user_id 
    where r.creation_date  is null

Answer 5

Joe Celko has a complete chapter on this in SQL for Smarties (calling it Runs and Sequences). I don't have that book at home, so when I get to work... I'll actually answer this. (assuming history table is called dbo.UserHistory and the number of days is @Days)

Another lead is from SQL Team's blog on runs

The other idea I've had, but don't have a SQL server handy to work on here is to use a CTE with a partitioned ROW_NUMBER like this:

WITH Runs
AS
  (SELECT UserID
         , CreationDate
         , ROW_NUMBER() OVER(PARTITION BY UserId
                             ORDER BY CreationDate)
           - ROW_NUMBER() OVER(PARTITION BY UserId, NoBreak
                               ORDER BY CreationDate) AS RunNumber
  FROM
     (SELECT UH.UserID
           , UH.CreationDate
           , ISNULL((SELECT TOP 1 1 
              FROM dbo.UserHistory AS Prior 
              WHERE Prior.UserId = UH.UserId 
              AND Prior.CreationDate
                  BETWEEN DATEADD(dd, DATEDIFF(dd, 0, UH.CreationDate), -1)
                  AND DATEADD(dd, DATEDIFF(dd, 0, UH.CreationDate), 0)), 0) AS NoBreak
      FROM dbo.UserHistory AS UH) AS Consecutive
)
SELECT UserID, MIN(CreationDate) AS RunStart, MAX(CreationDate) AS RunEnd
FROM Runs
GROUP BY UserID, RunNumber
HAVING DATEDIFF(dd, MIN(CreationDate), MAX(CreationDate)) >= @Days

The above is likely WAY HARDER than it has to be, but left as an a brain tickle for when you have some other definition of "a run" than just dates.

Answer 6

I used a simple math property to identify who consecutively accessed the site. This property is that you should have the day difference between the first time access and last time equal to number of records in your access table log.

Here are SQL script that I tested in Oracle DB (it should work in other DBs as well):

-- show basic understand of the math properties 
  select    ceil(max (creation_date) - min (creation_date))
              max_min_days_diff,
           count ( * ) real_day_count
    from   user_access_log
group by   user_id;


-- select all users that have consecutively accessed the site 
  select   user_id
    from   user_access_log
group by   user_id
  having       ceil(max (creation_date) - min (creation_date))
           / count ( * ) = 1;



-- get the count of all users that have consecutively accessed the site 
  select   count(user_id) user_count
    from   user_access_log
group by   user_id
  having   ceil(max (creation_date) - min (creation_date))
           / count ( * ) = 1;

Table prep script:

-- create table 
create table user_access_log (id           number, user_id      number, creation_date date);


-- insert seed data 
insert into user_access_log (id, user_id, creation_date)
  values   (1, 12, sysdate);

insert into user_access_log (id, user_id, creation_date)
  values   (2, 12, sysdate + 1);

insert into user_access_log (id, user_id, creation_date)
  values   (3, 12, sysdate + 2);

insert into user_access_log (id, user_id, creation_date)
  values   (4, 16, sysdate);

insert into user_access_log (id, user_id, creation_date)
  values   (5, 16, sysdate + 1);

insert into user_access_log (id, user_id, creation_date)
  values   (6, 16, sysdate + 5);