好的，这为您提供了一个索引视图（需要一个额外的视图来整理谁欠谁的详细信息），但它可能仍然不能满足您的要求。
/* Transactions table, as before, but with handy unique constraint for FK Target */
CREATE TABLE [dbo].[Transactions](
[Id] [int] IDENTITY(1,1) NOT NULL,
[Date] [datetime] NOT NULL,
[Amount] [float] NOT NULL,
[UserId] [bigint] NOT NULL,
[Remarks] [nvarchar](255) NULL,
[GroupFbGroupId] [bigint] NULL,
CONSTRAINT [PK_Transactions] PRIMARY KEY CLUSTERED (Id),
constraint UQ_Transactions_XRef UNIQUE (Id,Amount,UserId)
)
到目前为止没有什么令人惊讶的，我希望
/* Much expanded TransactionUser table, we'll hide it away and most of the maintenance is automatic */
CREATE TABLE [dbo]._TransactionUser(
[TransactionsPaidFor_Id] int NOT NULL,
[PaidForUsers_FbUserId] [bigint] NOT NULL,
Amount float not null,
PaidByUserId bigint not null,
UserCount int not null,
LowUserID as CASE WHEN [PaidForUsers_FbUserId] < PaidByUserId THEN [PaidForUsers_FbUserId] ELSE PaidByUserId END,
HighUserID as CASE WHEN [PaidForUsers_FbUserId] < PaidByUserId THEN PaidByUserId ELSE [PaidForUsers_FbUserId] END,
PerUserDelta as (Amount/UserCount) * CASE WHEN [PaidForUsers_FbUserId] < PaidByUserId THEN -1 ELSE 1 END,
constraint PK__TransactionUser PRIMARY KEY ([TransactionsPaidFor_Id],[PaidForUsers_FbUserId]),
constraint FK__TransactionUser_Transactions FOREIGN KEY ([TransactionsPaidFor_Id]) references dbo.Transactions,
constraint FK__TransactionUser_Transaction_XRef FOREIGN KEY ([TransactionsPaidFor_Id],Amount,PaidByUserID)
references dbo.Transactions (Id,Amount,UserId) ON UPDATE CASCADE
)
该表现在保留了足够的信息以允许构建视图。我们要做的其余工作是构建/维护表中的数据。请注意，通过外键约束，我们已经确保如果事务表中的金额发生更改，则所有内容都会重新计算。
/* View that mimics the original TransactionUser table -
in fact it has the same name so existing code doesn't need to change */
CREATE VIEW dbo.TransactionUser
with schemabinding
as
select
[TransactionsPaidFor_Id],
[PaidForUsers_FbUserId]
from
dbo._TransactionUser
GO
/* Effectively the PK on the original table */
CREATE UNIQUE CLUSTERED INDEX PK_TransactionUser on dbo.TransactionUser ([TransactionsPaidFor_Id],[PaidForUsers_FbUserId])
任何已经写出来反对的东西TransactionUser现在都将反对这种观点，而且不会更明智。除了，他们不能在没有帮助的情况下插入/更新/删除行：
/* Now we write the trigger that maintains the underlying table */
CREATE TRIGGER dbo.T_TransactionUser_IUD
ON dbo.TransactionUser
INSTEAD OF INSERT, UPDATE, DELETE
AS
SET NOCOUNT ON;
/* Every delete affects *every* row for the same transaction
We need to drop the counts on every remaining row, as well as removing the actual rows we're interested in */
WITH DropCounts as (
select TransactionsPaidFor_Id,COUNT(*) as Cnt from deleted group by TransactionsPaidFor_Id
), KeptRows as (
select tu.TransactionsPaidFor_Id,tu.PaidForUsers_FbUserId,UserCount - dc.Cnt as NewCount
from dbo._TransactionUser tu left join deleted d
on tu.TransactionsPaidFor_Id = d.TransactionsPaidFor_Id and
tu.PaidForUsers_FbUserId = d.PaidForUsers_FbUserId
inner join DropCounts dc
on
tu.TransactionsPaidFor_Id = dc.TransactionsPaidFor_Id
where
d.PaidForUsers_FbUserId is null
), ChangeSet as (
select TransactionsPaidFor_Id,PaidForUsers_FbUserId,NewCount,1 as Keep
from KeptRows
union all
select TransactionsPaidFor_Id,PaidForUsers_FbUserId,null,0
from deleted
)
merge into dbo._TransactionUser tu
using ChangeSet cs on tu.TransactionsPaidFor_Id = cs.TransactionsPaidFor_Id and tu.PaidForUsers_FbUserId = cs.PaidForUsers_FbUserId
when matched and cs.Keep = 1 then update set UserCount = cs.NewCount
when matched then delete;
/* Every insert affects *every* row for the same transaction
This is why the indexed view couldn't be generated */
WITH TU as (
select TransactionsPaidFor_Id,PaidForUsers_FbUserId,Amount,PaidByUserId from dbo._TransactionUser
where TransactionsPaidFor_Id in (select TransactionsPaidFor_Id from inserted)
union all
select TransactionsPaidFor_Id,PaidForUsers_FbUserId,Amount,UserId
from inserted i inner join dbo.Transactions t on i.TransactionsPaidFor_Id = t.Id
), CountedTU as (
select TransactionsPaidFor_Id,PaidForUsers_FbUserId,Amount,PaidByUserId,
COUNT(*) OVER (PARTITION BY TransactionsPaidFor_Id) as Cnt
from TU
)
merge into dbo._TransactionUser tu
using CountedTU new on tu.TransactionsPaidFor_Id = new.TransactionsPaidFor_Id and tu.PaidForUsers_FbUserId = new.PaidForUsers_FbUserId
when matched then update set Amount = new.Amount,PaidByUserId = new.PaidByUserId,UserCount = new.Cnt
when not matched then insert
([TransactionsPaidFor_Id],[PaidForUsers_FbUserId],Amount,PaidByUserId,UserCount)
values (new.TransactionsPaidFor_Id,new.PaidForUsers_FbUserId,new.Amount,new.PaidByUserId,new.Cnt);
现在正在维护基础表，我们终于可以首先编写您想要的索引视图......几乎。问题是我们创建的总数可能是正数或负数，因为我们已经对交易进行了标准化，以便我们可以轻松地将它们相加：
CREATE VIEW [dbo]._FriendBalances
WITH SCHEMABINDING
as
SELECT
LowUserID,
HighUserID,
SUM(PerUserDelta) as Balance,
COUNT_BIG(*) as Cnt
FROM dbo._TransactionUser
WHERE LowUserID != HighUserID
GROUP BY
LowUserID,
HighUserID
GO
create unique clustered index IX__FriendBalances on dbo._FriendBalances (LowUserID, HighUserID)
所以我们最终创建了一个视图，建立在上面的索引视图上，如果余额是负数，我们翻转欠债的人和欠债的人。但它将使用上述视图上的索引，这是我们试图通过使用索引视图来节省的大部分工作：
create view dbo.FriendBalances
as
select
CASE WHEN Balance >= 0 THEN LowUserID ELSE HighUserID END as PaidBy,
CASE WHEN Balance >= 0 THEN HighUserID ELSE LowUserID END as PaidFor,
ABS(Balance) as Balance
from
dbo._FriendBalances WITH (NOEXPAND)
现在，最后，我们插入您的示例数据：
set identity_insert dbo.Transactions on --Ensure we get IDs we know
GO
insert into dbo.Transactions (Id,[Date] , Amount , UserId , Remarks ,GroupFbGroupId)
select 1 ,'2001-01-01T00:00:00.000', 3000, 9990 ,'this is a test', NULL union all
select 2 ,'2001-01-01T00:00:00.000', 3000, 9990 ,'this is a test', NULL union all
select 3 ,'2001-01-01T00:00:00.000', 3000, 9991 ,'this is a test', NULL
GO
set identity_insert dbo.Transactions off
GO
insert into dbo.TransactionUser (TransactionsPaidFor_Id, PaidForUsers_FbUserId)
select 1, 9991 union all
select 1, 9992 union all
select 1, 9993 union all
select 2, 9990 union all
select 2, 9991 union all
select 2, 9992 union all
select 3, 9990 union all
select 3, 9993 union all
select 3, 9994
并查询最终视图：
select * from dbo.FriendBalances
PaidBy PaidFor Balance
9990 9991 1000
9990 9992 2000
9990 9993 1000
9991 9993 1000
9991 9994 1000
现在，我们可以做一些额外的工作，如果我们担心有人可能会找到一种方法来躲避触发器并直接更改基表。第一个是另一个索引视图，它将确保同一事务的每一行都具有相同 UserCount的值。最后，通过一些额外的列、检查约束、FK 约束和触发器中的更多工作，我认为我们可以确保它们UserCount是正确的——但它可能会增加比你想要的更多的开销。
如果您愿意，我可以为这些方面添加脚本 - 这取决于您希望/需要数据库的限制程度。