I have a table which logs the userid, course, sessionid and requestdate each time a webpage is loaded. I want to calcuate the duration per userid for a given courseid. It is problematic to do this due to overlapping timespans.
The data provided here should result in 10 minutes duration per user for course 1. I can't seem to get this right.
CREATE TABLE PageLogSample (
id INT NOT NULL PRIMARY KEY IDENTITY
, userid INT
, courseid INT
, sessionid INT
, requestdate DATETIME
);
TRUNCATE TABLE PageLogSample;
INSERT INTO PageLogSample (userid, courseid, sessionid, requestdate)
-- [0, 10] = 10 minutes
SELECT 1, 1, 1, '00:00:00'
UNION ALL SELECT 1, 1, 1, '00:10:00'
-- [0, 12] - [3, 5] = 10 minutes
-- or ... [0, 3] + [5, 12] = 10 minutes
UNION ALL SELECT 2, 1, 2, '00:00:00'
UNION ALL SELECT 2, 2, 2, '00:03:00'
UNION ALL SELECT 2, 2, 2, '00:05:00'
UNION ALL SELECT 2, 1, 2, '00:12:00'
-- [0, 12] - [3, 5] = 10 minutes
-- or ... [0, 3] + [5, 12] = 10 minutes
UNION ALL SELECT 3, 1, 3, '00:00:00'
UNION ALL SELECT 3, 2, 3, '00:03:00'
UNION ALL SELECT 3, 2, 3, '00:05:00'
UNION ALL SELECT 3, 1, 3, '00:12:00'
UNION ALL SELECT 3, 2, 3, '00:15:00'
-- [1, 13] - [3, 5] = 10 minutes
-- or ... [1, 3] + [5, 13] = 10 minutes
UNION ALL SELECT 4, 2, 4, '00:00:00'
UNION ALL SELECT 4, 1, 4, '00:01:00'
UNION ALL SELECT 4, 2, 4, '00:03:00'
UNION ALL SELECT 4, 2, 4, '00:05:00'
UNION ALL SELECT 4, 1, 4, '00:13:00'
UNION ALL SELECT 4, 2, 4, '00:15:00'
-- [0, 5] + [10, 15] = 10 minutes
UNION ALL SELECT 5, 1, 5, '00:00:00'
UNION ALL SELECT 5, 1, 5, '00:05:00'
UNION ALL SELECT 5, 1, 6, '00:10:00'
UNION ALL SELECT 5, 1, 6, '00:15:00'
-- [0, 10] = 10 minutes (ignoring everything inbetween)
UNION ALL SELECT 6, 1, 7, '00:00:00'
UNION ALL SELECT 6, 1, 7, '00:03:00'
UNION ALL SELECT 6, 1, 7, '00:05:00'
UNION ALL SELECT 6, 1, 7, '00:07:00'
UNION ALL SELECT 6, 1, 7, '00:10:00'
-- [0, 11] - [5, 6] = 10 minutes
-- or ... [0, 3] + [7, 11] = 6 minutes (good)
-- or ... [0, 5] + [7, 11] = 9 minutes (better)
UNION ALL SELECT 7, 1, 8, '00:00:00'
UNION ALL SELECT 7, 1, 8, '00:03:00'
UNION ALL SELECT 7, 2, 8, '00:05:00'
UNION ALL SELECT 7, 2, 8, '00:06:00'
UNION ALL SELECT 7, 1, 8, '00:07:00'
UNION ALL SELECT 7, 1, 8, '00:11:00'
-- [0, 1] + [2, 4] + [5, 7] + [8, 13] = 10
UNION ALL SELECT 8, 1, 9, '00:00:00'
UNION ALL SELECT 8, 2, 9, '00:01:00'
UNION ALL SELECT 8, 1, 9, '00:02:00'
UNION ALL SELECT 8, 1, 9, '00:03:00'
UNION ALL SELECT 8, 2, 9, '00:04:00'
UNION ALL SELECT 8, 1, 9, '00:05:00'
UNION ALL SELECT 8, 1, 9, '00:06:00'
UNION ALL SELECT 8, 2, 9, '00:07:00'
UNION ALL SELECT 8, 1, 9, '00:08:00'
UNION ALL SELECT 8, 1, 9, '00:13:00'
;
first trying the naive approach. This gives mistakes with overlapping parts of the session.
DECLARE @courseid INT;
SET @courseid = 1;
SELECT subquery.userid
, COUNT(DISTINCT subquery.sessionid) AS sessioncount
, SUM(subquery.duration) AS duration
, CASE SUM(subquery.duration)
WHEN 10 THEN 'ok'
ELSE 'ERROR'
END
FROM (
SELECT userid
, sessionid
, DATEDIFF(MINUTE, MIN(requestdate), MAX(requestdate)) AS duration
FROM PageLogSample
WHERE courseid = @courseid
GROUP BY userid
, sessionid
) subquery
GROUP BY subquery.userid
ORDER BY subquery.userid;
-- userid sessioncount duration
-- 1 1 10 ok
-- 2 1 12 ERROR
-- 3 1 12 ERROR
-- 4 1 12 ERROR
-- 5 2 10 ok
Second try. Avoid overlappings. This only works partially.
DECLARE @courseid INT;
SET @courseid = 1;
WITH cte (userid, courseid, sessionid, start, finish, duration)
AS (
SELECT userid
, courseid
, sessionid
, MIN(requestdate)
, MAX(requestdate)
, DATEDIFF(MINUTE, MIN(requestdate), MAX(requestdate))
FROM P开发者_如何学编程ageLogSample
GROUP BY userid
, courseid
, sessionid
)
SELECT naive.userid
, naive.sessioncount
, naive.duration AS naiveduration
, correction.duration AS correctionduration
, naive.duration - ISNULL(correction.duration, 0) AS duration
, CASE naive.duration - ISNULL(correction.duration, 0)
WHEN 10 THEN 'ok'
ELSE 'ERROR'
END
FROM (
SELECT cte.userid
, COUNT(DISTINCT cte.sessionid) AS sessioncount
, SUM(cte.duration) AS duration
FROM cte
WHERE cte.courseid = @courseid
GROUP BY cte.userid
) naive
LEFT JOIN (
SELECT errors.userid
, SUM(errors.duration) AS duration
FROM cte errors
WHERE errors.courseid <> @courseid
AND EXISTS (
SELECT *
FROM cte
WHERE cte.start <= errors.start
AND cte.finish >= errors.finish
AND cte.courseid = @courseid
)
GROUP BY errors.userid
) correction
ON naive.userid = correction.userid
;
-- userid sessioncount naiveduration correctionduration duration
-- 1 1 10 NULL 10 ok
-- 2 1 12 2 10 ok
-- 3 1 12 NULL 12 ERROR
-- 4 1 12 NULL 12 ERROR
-- 5 2 10 NULL 10 ok
Update: Ed Harpers comment Really made me rethink my approach.
So here comes the third trial. Here I first search for which rows represent an entrance into the course and which represent someone leaving. Then I take the sum of all endtimes and substract the sum of all begintimes. I think it is more correct, while not perfect.
DECLARE @courseid INT;
SET @courseid = 1;
WITH numberedcte (rn, id, userid, courseid, sessionid, requestdate)
AS (
SELECT ROW_NUMBER() OVER (PARTITION BY sessionid, userid ORDER BY id)
, id
, userid
, courseid
, sessionid
, requestdate
FROM PageLogSample
)
, typedcte (rowtype, id, userid, courseid, sessionid, requestdate, nextrequestdate)
AS (
SELECT CASE
WHEN previousrequest.courseid = nextrequest.courseid
THEN 'between'
WHEN previousrequest.courseid IS NULL
OR nextrequest.courseid = numberedcte.courseid
THEN 'begin'
WHEN nextrequest.courseid IS NULL
OR previousrequest.courseid = numberedcte.courseid
THEN 'end'
ELSE 'error?'
END AS rowtype
, numberedcte.id
, numberedcte.userid
, numberedcte.courseid
, numberedcte.sessionid
, numberedcte.requestdate
, nextrequest.requestdate
FROM numberedcte
LEFT JOIN numberedcte previousrequest
ON previousrequest.userid = numberedcte.userid
AND previousrequest.sessionid = numberedcte.sessionid
AND previousrequest.rn = numberedcte.rn - 1
LEFT JOIN numberedcte nextrequest
ON nextrequest.userid = numberedcte.userid
AND nextrequest.sessionid = numberedcte.sessionid
AND nextrequest.rn = numberedcte.rn + 1
WHERE numberedcte.courseid = @courseid
AND (
nextrequest.courseid = @courseid
OR previousrequest.courseid = @courseid
)
)
, beginsum (userid, value)
AS (
SELECT userid, SUM(DATEPART(MINUTE, requestdate))
FROM typedcte
WHERE rowtype = 'begin'
GROUP BY userid
)
, endsum (userid, value)
AS (
SELECT userid, SUM(DATEPART(MINUTE, ISNULL(nextrequestdate, requestdate)))
FROM typedcte
WHERE rowtype = 'end'
GROUP BY userid
)
SELECT beginsum.userid
, endsum.value - beginsum.value AS duration
FROM beginsum
INNER JOIN endsum
ON beginsum.userid = endsum.userid
;
The only problem here is that I only get output for user 1 and 5 from my original sample data. The added user 6 also gives correct output. The added user 7 gives me a satisfactory output now. User 8 is almost perfect, I miss one minute from the first row to the second.
-- userid duration
-- 1 10
-- 5 10
-- 6 10
-- 7 9
-- 8 9
I feel like I'm inches away from getting this completely right. The only durations missing are from the pagerequests that didn't happen in groups. Can someone help me find a way to get the lonely pageviews?
Update: Here comes a fourth trial. Here I assign a value to each request and sum them up. It doesn't give me exactlu the output I hoped for, but looks like it could be good enough.
DECLARE @courseid INT;
SET @courseid = 1;
WITH numberedcte (rn, userid, courseid, sessionid, requestdate)
AS (
SELECT ROW_NUMBER() OVER (PARTITION BY sessionid, userid ORDER BY id)
, userid
, courseid
, sessionid
, requestdate
FROM PageLogSample
)
, valuecte (value, userid, courseid, sessionid)
AS (
SELECT CASE
--alone
WHEN ( previousrequest.courseid IS NULL
OR previousrequest.courseid <> numberedcte.courseid
)
AND nextrequest.courseid <> numberedcte.courseid
THEN DATEDIFF(MINUTE, numberedcte.requestdate, nextrequest.requestdate)
--between
WHEN previousrequest.courseid = nextrequest.courseid
THEN 0
--begin
WHEN previousrequest.courseid IS NULL
OR nextrequest.courseid = numberedcte.courseid
THEN -1 * DATEPART(MINUTE, numberedcte.requestdate)
--ignored (end with no next request)
WHEN nextrequest.courseid IS NULL
AND previousrequest.courseid <> numberedcte.courseid
THEN 0
--end
WHEN nextrequest.courseid IS NULL
OR previousrequest.courseid = numberedcte.courseid
THEN DATEPART(MINUTE, ISNULL(nextrequest.requestdate, numberedcte.requestdate))
--impossible?
ELSE 0
END
, numberedcte.userid
, numberedcte.courseid
, numberedcte.sessionid
FROM numberedcte
LEFT JOIN numberedcte previousrequest
ON previousrequest.userid = numberedcte.userid
AND previousrequest.sessionid = numberedcte.sessionid
AND previousrequest.rn = numberedcte.rn - 1
LEFT JOIN numberedcte nextrequest
ON nextrequest.userid = numberedcte.userid
AND nextrequest.sessionid = numberedcte.sessionid
AND nextrequest.rn = numberedcte.rn + 1
WHERE numberedcte.courseid = @courseid
)
SELECT userid
, courseid
, COUNT(DISTINCT sessionid) AS sessioncount
, SUM(value) AS duration
FROM valuecte
GROUP BY userid
, courseid
ORDER BY userid
;
As you can see the results are not entirely what I expected.
-- userid courseid sessioncount duration
-- 1 1 1 10
-- 2 1 1 3
-- 3 1 1 6
-- 4 1 1 4
-- 5 1 2 10
-- 6 1 1 10
-- 7 1 1 9
-- 8 1 1 10
Performance is horrible on my local copy of the real database. So if anyone has ideas as to write this in a more performant way ... shoot.
Update: Performance is up. I added an index and it works a charm now.
Sorry but I think you have a data problem. Looking at the sample data provided user 2 is in courseid 1 for 12 minutes and courseid 2 for 2 minutes.
Are you sure you have supplied the correct data?
This is as close as I can get. It fails for userid 4.
As I said in my comment, requestdate
is sometimes a start and sometimes an end of a course, and I can't see a simple general rule for deriving which role it plays on a given row.
DECLARE @courseid INT;
SET @courseid = 1;
WITH orderCTE
AS
(
SELECT *
,ROW_NUMBER() OVER (PARTITION BY sessionid
ORDER BY id
) AS rn
FROM PageLogSample
--order by rn
)
,startendCTE
AS
(
SELECT CASE WHEN start1.rn = 1
THEN start1.courseid
ELSE end1.courseid
END courseid
,start1.sessionid
,start1.userid
,DATEDIFF(mi,start1.requestdate,end1.requestdate) duration
FROM orderCTE AS start1
JOIN orderCTE AS end1
ON end1.rn = start1.rn + 1
AND end1.sessionid = start1.sessionid
)
SELECT courseid
,COUNT(1) sessionCount
,userid
,SUM(duration) totalDuration
FROM startendCTE
WHERE courseid = @courseid
GROUP BY courseid
,userid;
This is pretty messy, but it appears to be working for CourseID 1. I didn't try it with other courses, so you may want to test that! :D
The basic premise is that I'm getting the duration of time between the first and last session of the target CourseID and then I'm subtracting the duration of any sessions that were not of the specified CourseID, but where the session request time fell within the min and max request times of the targeted CourseID. I hope that makes sense.
The Query could definitely be cleaned up, possibly with a CTE or something. Interesting question BTW! :)
DECLARE @courseid INT;
SET @courseid = 1;
SELECT
TargetCourse.UserID,
COUNT(Distinct(TargetCourse.SessionID)) as SessionCount,
SUM(TargetCourse.Duration - Coalesce(OtherCourses.Duration,0)) as Duration
FROM
(
SELECT
TargetCourse.UserID, TargetCourse.SessionID,
MIN(TargetCourse.RequestDate) FirstRequest, MAX(TargetCourse.RequestDate) LastRequest,
DATEDIFF(MINUTE, MIN(TargetCourse.RequestDate), MAX(TargetCourse.RequestDate)) AS duration
FROM
PageLogSample TargetCourse
WHERE
TargetCourse.CourseID = @courseid
GROUP BY
TargetCourse.UserID, TargetCourse.SessionID
) as TargetCourse
LEFT OUTER JOIN
(
SELECT
OtherCourses.UserID, OtherCourses.SessionID,
MIN(OtherCourses.RequestDate) AS FirstRequest, MAX(OtherCourses.RequestDate) AS LastRequest,
DATEDIFF(MINUTE, MIN(OtherCourses.RequestDate), MAX(OtherCourses.RequestDate)) AS duration
FROM
PageLogSample OtherCourses
WHERE
OtherCourses.CourseID <> @courseid AND
OtherCourses.RequestDate between
(Select MIN(RequestDate) From PageLogSample T Where T.UserID = OtherCourses.UserID and T.CourseID = @courseid) AND
(Select MAX(RequestDate) From PageLogSample T Where T.UserID = OtherCourses.UserID and T.CourseID = @courseid)
GROUP BY
OtherCourses.UserID, OtherCourses.SessionID
) as OtherCourses ON
OtherCourses.UserID = TargetCourse.UserID AND
OtherCourses.FirstRequest BETWEEN TargetCourse.FirstRequest and TargetCourse.LastRequest
Group By TargetCourse.UserID
Some more sample data and a hopefully logical assumption of how much time each user spent in each course.
INSERT INTO PageLogSample (userid, courseid, sessionid, requestdate)
-- [0, 10] = 10 minutes
SELECT 1, 1, 1, '00:00:00'
UNION ALL SELECT 1, 1, 1, '00:10:00'
-- [0, 3] = 3 minutes
-- there is no way to know how long the user was on that last page
UNION ALL SELECT 2, 1, 2, '00:00:00'
UNION ALL SELECT 2, 2, 2, '00:03:00'
UNION ALL SELECT 2, 2, 2, '00:05:00'
UNION ALL SELECT 2, 1, 2, '00:12:00'
-- [0, 3] + [12, 15] = 6 minutes
-- the [5, 12] part was spent on a page of course 2
UNION ALL SELECT 3, 1, 3, '00:00:00'
UNION ALL SELECT 3, 2, 3, '00:03:00'
UNION ALL SELECT 3, 2, 3, '00:05:00'
UNION ALL SELECT 3, 1, 3, '00:12:00'
UNION ALL SELECT 3, 2, 3, '00:15:00'
-- [1, 3] + [13, 15] = 4 minutes
UNION ALL SELECT 4, 2, 4, '00:00:00'
UNION ALL SELECT 4, 1, 4, '00:01:00'
UNION ALL SELECT 4, 2, 4, '00:03:00'
UNION ALL SELECT 4, 2, 4, '00:05:00'
UNION ALL SELECT 4, 1, 4, '00:13:00'
UNION ALL SELECT 4, 2, 4, '00:15:00'
-- [0, 5] + [10, 15] = 10 minutes
UNION ALL SELECT 5, 1, 5, '00:00:00'
UNION ALL SELECT 5, 1, 5, '00:05:00'
UNION ALL SELECT 5, 1, 6, '00:10:00'
UNION ALL SELECT 5, 1, 6, '00:15:00'
-- [0, 10] = 10 minutes (ignoring everything inbetween)
UNION ALL SELECT 6, 1, 7, '00:00:00'
UNION ALL SELECT 6, 1, 7, '00:03:00'
UNION ALL SELECT 6, 1, 7, '00:05:00'
UNION ALL SELECT 6, 1, 7, '00:07:00'
UNION ALL SELECT 6, 1, 7, '00:10:00'
-- [0, 5] + [7, 11] = 9 minutes
UNION ALL SELECT 7, 1, 8, '00:00:00'
UNION ALL SELECT 7, 1, 8, '00:03:00'
UNION ALL SELECT 7, 2, 8, '00:05:00'
UNION ALL SELECT 7, 2, 8, '00:06:00'
UNION ALL SELECT 7, 1, 8, '00:07:00'
UNION ALL SELECT 7, 1, 8, '00:11:00'
-- [0, 1] + [2, 4] + [5, 7] + [8, 13] = 10
UNION ALL SELECT 8, 1, 9, '00:00:00'
UNION ALL SELECT 8, 2, 9, '00:01:00'
UNION ALL SELECT 8, 1, 9, '00:02:00'
UNION ALL SELECT 8, 1, 9, '00:03:00'
UNION ALL SELECT 8, 2, 9, '00:04:00'
UNION ALL SELECT 8, 1, 9, '00:05:00'
UNION ALL SELECT 8, 1, 9, '00:06:00'
UNION ALL SELECT 8, 2, 9, '00:07:00'
UNION ALL SELECT 8, 1, 9, '00:08:00'
UNION ALL SELECT 8, 1, 9, '00:13:00'
-- there is nothing we can say about either of there requests
-- 0 minutes
UNION ALL SELECT 9, 1, 10, '00:10:00'
UNION ALL SELECT 9, 1, 11, '00:20:00'
;
Now we get our data like this:
WITH numberedcte (rn, userid, courseid, sessionid, requestdate)
AS (
SELECT ROW_NUMBER() OVER (PARTITION BY sessionid, userid ORDER BY id)
, userid
, courseid
, sessionid
, requestdate
FROM PageLogSample
)
, valuecte (value, userid, courseid, sessionid)
AS (
SELECT CASE
--alone in session
WHEN previousrequest.courseid IS NULL
AND nextrequest.courseid IS NULL
THEN 0
--alone
WHEN ( previousrequest.courseid IS NULL
OR previousrequest.courseid <> numberedcte.courseid
)
AND nextrequest.courseid <> numberedcte.courseid
THEN DATEDIFF(MINUTE, numberedcte.requestdate, nextrequest.requestdate)
--between
WHEN previousrequest.courseid = nextrequest.courseid
THEN 0
--begin
WHEN previousrequest.courseid IS NULL
OR nextrequest.courseid = numberedcte.courseid
THEN -1 * DATEPART(MINUTE, numberedcte.requestdate)
--ignored (end with no next request)
WHEN nextrequest.courseid IS NULL
AND previousrequest.courseid <> numberedcte.courseid
THEN 0
--end
WHEN nextrequest.courseid IS NULL
OR previousrequest.courseid = numberedcte.courseid
THEN DATEPART(MINUTE, ISNULL(nextrequest.requestdate, numberedcte.requestdate))
--impossible?
ELSE 0
END
, numberedcte.userid
, numberedcte.courseid
, numberedcte.sessionid
FROM numberedcte
LEFT JOIN numberedcte previousrequest
ON previousrequest.userid = numberedcte.userid
AND previousrequest.sessionid = numberedcte.sessionid
AND previousrequest.rn = numberedcte.rn - 1
LEFT JOIN numberedcte nextrequest
ON nextrequest.userid = numberedcte.userid
AND nextrequest.sessionid = numberedcte.sessionid
AND nextrequest.rn = numberedcte.rn + 1
WHERE numberedcte.courseid = @courseid
)
SELECT userid
, courseid
, COUNT(DISTINCT sessionid) AS sessioncount
, SUM(value) AS duration
FROM valuecte
GROUP BY userid
, courseid
ORDER BY userid
;
This is the result I get. I'm quite pleased with it. Notice how the session count remains correct for user 9.
userid courseid sessioncount duration
1 1 1 10
2 1 1 3
3 1 1 6
4 1 1 4
5 1 2 10
6 1 1 10
7 1 1 9
8 1 1 10
9 1 2 0
"The data is correct, but is is just difficult to get relevant meaning out of it."
I'm pressed to respond that this a contradiction of terms. Data of which you don't know what it means is not data.
As for your original question :
What you need is a DBMS that offers decent support for INTERVAL types. No SQL system plays in that league. Apart from a few tutorial systems, my own DBMS (no further pushing that in this context, so no link) is the only one I know of that offers the kind of support that is really needed for such problems.
If you're interested, google around for "interval types", "packed normal form", "temporal data" and you'll run into it eventually.
精彩评论