SQL Table-value function optimisation / improvement - sql

Question

We've got a query that is taking a very long time to complete with a large dataset. I think I've tracked it down to a table-value function in the SQL server.
The query is designed to return the difference in printing usage between two dates. So if a printer had usage of 100 at date x and 200 at date y a row needs to be returned which reflects that it has had a usage change of 100.
These readings are taken periodically (but not every day) and stored in a table called MeterReadings. The code for the table-value function is below. This is then called from another SQL query which joins the returned table on a devices table with an inner join to get extra device information.
Any advise as to how to optimise the below would be appreciated.
ALTER FUNCTION [dbo].[DeviceUsage]
-- Add the parameters for the stored procedure here
( #StartDate DateTime , #EndDate DateTime )
RETURNS table
AS
RETURN
(
SELECT MAX(dbo.MeterReadings.ScanDateTime) AS MX,
MAX(dbo.MeterReadings.DeviceTotal - reading.DeviceTotal) AS TotalDiff,
MAX(dbo.MeterReadings.TotalCopy - reading.TotalCopy) AS CopyDiff,
MAX(dbo.MeterReadings.TotalPrint - reading.TotalPrint) AS PrintDiff,
MAX(dbo.MeterReadings.TotalScan - reading.TotalScan) AS ScanDiff,
MAX(dbo.MeterReadings.TotalFax - reading.TotalFax) AS FaxDiff,
MAX(dbo.MeterReadings.TotalMono - reading.TotalMono) AS MonoDiff,
MAX(dbo.MeterReadings.TotalColour - reading.TotalColour) AS ColourDiff,
MIN(reading.ScanDateTime) AS MN, dbo.MeterReadings.DeviceID
FROM dbo.MeterReadings INNER JOIN (SELECT * FROM dbo.MeterReadings WHERE
(dbo.MeterReadings.ScanDateTime > #StartDate) AND
(dbo.MeterReadings.ScanDateTime < #EndDate) )
AS reading ON dbo.MeterReadings.DeviceID = reading.DeviceID
WHERE (dbo.MeterReadings.ScanDateTime > #StartDate) AND (dbo.MeterReadings.ScanDateTime < #EndDate)
GROUP BY dbo.MeterReadings.DeviceID);

Answer 1

On the assumption that a value can only ever increase over time, it can certainly be simplified.
SELECT
DeviceID,
MIN(ScanDateTime) AS MN,
MAX(ScanDateTime) AS MX,
MAX(DeviceTotal ) - MIN(DeviceTotal) AS TotalDiff,
MAX(TotalCopy ) - MIN(TotalCopy ) AS CopyDiff,
MAX(TotalPrint ) - MIN(TotalPrint ) AS PrintDiff,
MAX(TotalScan ) - MIN(TotalScan ) AS ScanDiff,
MAX(TotalFax ) - MIN(TotalFax ) AS FaxDiff,
MAX(TotalMono ) - MIN(TotalMono ) AS MonoDiff,
MAX(TotalColour ) - MIN(TotalColour) AS ColourDiff
FROM
dbo.MeterReadings
WHERE
ScanDateTime > #StartDate
AND ScanDateTime < #EndDate
GROUP BY
DeviceID
This assumes that if you have reading on dates 1, 3, 5, 7, 9 and you want to report on 2 -> 8 then you want reading 7 - reading 3. I would have thought you wanted reading 7 - reading 1?
The above query should be fine for relatively small ranges. If you have Huge ranges of time, the MAX() - MIN() will be operating on large numbers of rows. This can then possibly be improved even further with the following (with correlated sub-queries to lookup just the two rows that you want).
As a side benefit, this also works even if the values can go down as well as up.
(I assume the existance of a Device table for a simpler query and faster performance.)
SELECT
Device.DeviceID,
start.ScanDateTime AS MN,
finish.ScanDateTime AS MX,
finish.DeviceTotal - start.DeviceTotal AS TotalDiff,
finish.TotalCopy - start.TotalCopy AS CopyDiff,
finish.TotalPrint - start.TotalPrint AS PrintDiff,
finish.TotalScan - start.TotalScan AS ScanDiff,
finish.TotalFax - start.TotalFax AS FaxDiff,
finish.TotalMono - start.TotalMono AS MonoDiff,
finish.TotalColour - start.TotalColour AS ColourDiff
FROM
dbo.Device AS device
INNER JOIN
dbo.MeterReadings AS start
ON start.DeviceID = device.DeviceID
AND start.ScanDateTime = (SELECT MIN(ScanDateTime)
FROM dbo.MeterReadings
WHERE DeviceID = device.DeviceID
AND ScanDateTime > #startDate
AND ScanDateTime < #endDate)
INNER JOIN
dbo.MeterReadings AS finish
ON finish.DeviceID = device.DeviceID
AND finish.ScanDateTime = (SELECT MAX(ScanDateTime)
FROM dbo.MeterReadings
WHERE DeviceID = device.DeviceID
AND ScanDateTime > #startDate
AND ScanDateTime < #endDate)
This can also be modified to pick up the start as being the first date on or before #startDate, if required.
EDIT: Modification to pick the start reading as being for the first date on or before #startDate.
SELECT
Device.DeviceID,
start.ScanDateTime AS MN,
finish.ScanDateTime AS MX,
COALESCE(finish.DeviceTotal, 0) - COALESCE(start.DeviceTotal, 0) AS TotalDiff,
COALESCE(finish.TotalCopy , 0) - COALESCE(start.TotalCopy , 0) AS CopyDiff,
COALESCE(finish.TotalPrint , 0) - COALESCE(start.TotalPrint , 0) AS PrintDiff,
COALESCE(finish.TotalScan , 0) - COALESCE(start.TotalScan , 0) AS ScanDiff,
COALESCE(finish.TotalFax , 0) - COALESCE(start.TotalFax , 0) AS FaxDiff,
COALESCE(finish.TotalMono , 0) - COALESCE(start.TotalMono , 0) AS MonoDiff,
COALESCE(finish.TotalColour, 0) - COALESCE(start.TotalColour, 0) AS ColourDiff
FROM
dbo.Device AS device
LEFT JOIN
dbo.MeterReadings AS start
ON start.DeviceID = device.DeviceID
AND start.ScanDateTime = (SELECT MAX(ScanDateTime)
FROM dbo.MeterReadings
WHERE DeviceID = device.DeviceID
AND ScanDateTime < #startDate)
LEFT JOIN
dbo.MeterReadings AS finish
ON finish.DeviceID = device.DeviceID
AND finish.ScanDateTime = (SELECT MAX(ScanDateTime)
FROM dbo.MeterReadings
WHERE DeviceID = device.DeviceID
AND ScanDateTime < #endDate)

Answer 2

Your query seems to compute a cross-product of all readings in a time range for each particular device. This works semantically because the MIN and MAX aggregates don't care about duplicates. But this is very slow. If you are comparing 100 dates with themselves you need to process 10,000 rows.
I suggest you calculate the MIN and MAX values for each metric/column over the entire time interval and then subtract them. That way you don't need to join and you need a single pass ofer the data. Like this:
select Diff = MAX(col) - MIN(col)
from readings
group by DeviceID