SQL Server CREATE AGGREGATE logic error - sql-server

Question

I've created a CLR funtion for SQL Server 2014 that should calculate subtraction between the first and the last value in column [Value].
Here is the table:
Date_Time Value
-------------------------------------
2018-03-29 09:30:02.533 6771
2018-03-29 10:26:23.557 6779
2018-03-29 13:12:04.550 6787
2018-03-29 13:55:44.560 6795
Here is the code:
using System;
using System.Data;
using System.Data.SqlClient;
using System.Data.SqlTypes;
using Microsoft.SqlServer.Server;
[Serializable]
[SqlUserDefinedAggregate(Format.Native,
IsInvariantToDuplicates = false,
IsInvariantToNulls = true,
IsInvariantToOrder = false,
IsNullIfEmpty = true,
Name = "SUBT")]
public struct SUBT
{
private double first;
private double last;
private int count;
public void Init()
{
first = 0.0;
last = 0.0;
count = 0;
}
public void Accumulate(SqlDouble Value)
{
if (!Value.IsNull)
{
if (count == 0)
first = (double)Value;
else
last = (double)Value;
count += 1;
}
}
public void Merge(SUBT Group)
{
first = Group.first;
last = Group.last;
count += 1;
}
public SqlDouble Terminate()
{
double value = (double)(last - first);
return new SqlDouble(value);
}
}
So the result should be [Value]=24, i.e. 6795 - 6771, but I get 6795 :(
Where is the error?

Answer 1

This aggregate function is dependent on order but there is no ordering guarantee for the aggregation input stream. Consequently, the results are execution plan dependent.
Assuming the Date_Time value is the desired ordering, you could provide both Date_Time and Value as function arguments, save the value with the lowest and highest Date_time values and use those in the Merge and Terminate methods.

Answer 2

There are a couple of problems here, I believe:
In the Merge method you are making two bad assumptions:
You are assuming that the incoming Group has a value, yet it could have been called on one or more NULL values, in which case all 3 internal variables are 0. Yet you are overwriting the current values of first and last, which could have non-0 values prior to Merge being called, but then will end up back at 0 due to being overwritten.
You are assuming that at least one of the instances — the current one or the incoming Group — has values set (i.e. has been called at least once on a non-NULL value). In the case that both instances have only been called with a NULL value, you will have 0 for first and last, yet you will increment counter. I am not sure if Accumulate will be called again once things are being merged, but if it does, you will skip setting first. This is not your problem at the moment since you do not have multiple (or any) NULL values, but it is a potential problem for real data sets.
In the case where both instances have been called on non-NULL values, both will at least have first set, and maybe last (or maybe not). By overwriting the current instance with the incoming Group, you could be losing the real first value, losing the real last value.
As #DanGuzman mentioned in his answer, there is no guaranteed ordering of User-Defined Aggregates (the IsInvariantToOrder property of the SqlUserDefinedAggregate attribute is ignored / unused). And, as he noted, you will need to pass in the Date_Time value in order to handle this aspect of the operation manually. However, it won't be used in the Terminate method. It will instead be used to compare to two new variables: firstDate and lastDate, initialized to a future and past, respectively (this will likely require changing the Format to UserDefined and then adding custom Read and Write methods -- unless you can store the full DateTime values as ticks, perhaps).
Get rid of the counter variable
In the Accumulate method, you will need to:
IF the incoming Date_Value against firstDate. If it is before to firstDate then store this new value as firstDate and store Value as first, ELSE
IF the incoming Date_Value against lastDate. If it is after to lastDate then store this new value as lastDate and store Value as last, ELSE do nothing
In the Merge method, do a similar comparison for firstDate between both instances and keep the earlier one (date and value). Do the same with lastDate and keep the later one (date and value). (Note: these changes should fix all of the Merge issues noted above in # 1)
The terminate method shouldn't change
For what it's worth, I ran the code exactly as you have posted in the question, and it returns the expected value, using the following test query:
CREATE TABLE #Test ([Date_Time] DATETIME, [Value] FLOAT);
-- TRUNCATE TABLE #Test;
INSERT INTO #Test VALUES ('2018-03-29 09:30:02.533', 6771);
INSERT INTO #Test VALUES ('2018-03-29 10:26:23.557', 6779);
INSERT INTO #Test VALUES ('2018-03-29 13:12:04.550', 6787);
INSERT INTO #Test VALUES ('2018-03-29 13:55:44.560', 6795);
SELECT dbo.SUBT([Value])
FROM #Test;
-- 24
So, if you are still having issues, then you will need to post more info, such as the test query (and maybe table) that you are using. But even if it appears to work, like it does on my system, it still has the potential ordering problem and will need to be updated as noted above regardless.
Other notes:
In the Accumulate method you have (double)Value. There is no need to cast the incoming parameter. All Sql* types have a Value property that returns the value in the native .NET type. In this case, just use Value.Value. That isn't great for readability, so consider changing the name of the input paramter ;-).
You never use the value of counter, so why increment it? You could instead just use a bool, and set it to true here. Setting it to true for each non-NULL value will not change the operation. However, this is a moot point since you truly need to set either first or last in each call to this UDA, based on the current Date_Time value.