I found a SQL deadlock scenario in my application during concurrency. I belive that the two statements that cause the deadlock are (note - I'm using LINQ2SQL and DataContext.ExecuteCommand(), that's where this.studioId.ToString() comes into play):
exec sp_executesql N'INSERT INTO HQ.dbo.SynchronizingRows ([StudioId], [UpdatedRowId])
SELECT @p0, [t0].[Id] FROM [dbo].[UpdatedRows] AS [t0] WHERE NOT (EXISTS(
SELECT NULL AS [EMPTY] FROM [dbo].[ReceivedUpdatedRows] AS [t1] WHERE ([t1].[StudioId] = @p0)
AND ([t1].[UpdatedRowId] = [t0].[Id]) ))',N'@p0 uniqueidentifier',@p0='" + this.studioId.ToString() + "';
and
exec sp_executesql N'INSERT INTO HQ.dbo.ReceivedUpdatedRows ([UpdatedRowId], [StudioId], [ReceiveDateTime])
SELECT [t0].[UpdatedRowId], @p0, GETDATE() FROM [dbo].[SynchronizingRows] AS [t0]
WHERE ([t0].[StudioId] = @p0)',N'@p0 uniqueidentifier',@p0='" + this.studioId.ToString() + "';
The basic logic of my (client-server) application is this:
Every time someone inserts or updates a row on the server side, I also insert a row into the table UpdatedRows, specifying the RowId of the modified row.
When a client tries to synchronize data, it first copies all of the rows in the UpdatedRows table, that don't contain a reference row for the specific client in the table ReceivedUpdatedRows, to the table SynchronizingRows (the first statement taking part in the deadlock). Afterwards, during the synchronization I look for modified rows via lookup of the SynchronizingRows table. This step is required, otherwise if someone inserts new rows or modifies rows on the server side during synchronization I will miss them and won't get them during the next synchronization (explanation scenario to long to write here...).
Once synchronization is complete, I insert rows to the ReceivedUpdatedRows table specifying that this client has received the UpdatedRows contained in the SynchronizingRows table (the second statement taking part in the deadlock).
Finally I delete all rows from the SynchronizingRows table that belong to the current client.
The way I see it, the deadlock is occuring on tables SynchronizingRows (abbreviation SR) and ReceivedUpdatedRows (abbreviation RUR) during steps 2 and 3 (one client is in step 2 and is inserting into SR and selecting from RUR; while another client is in step 3 inserting into RUR and selecting from SR).
I googled a bit about SQL deadlocks and came to a conclusion that I have three options. Inorder to make a decision I need more input about each option/workaround:
Workaround 1:
The first advice given on the web about SQL deadlocks - restructure tables/queries so that deadlocks don't happen in the first place. Only problem with this is that with my IQ I don't see a way to do the synchronization logic any differently. If someone wishes to dwelve deeper into my current synchronization logic, how and why it is set up the way it is, I'll post a link for the explanation. Perhaps, with the help of someone smarter than me, it's possible to create a logic that is deadlock free.
Workaround 2:
The second most common advice seems to be the use of WITH(NOLOCK) hint. The problem with this is that NOLOCK might miss or duplicate some rows. Duplication is not a problem, but missing rows is catastrophic! Another option is the WITH(READPAST) hint. On the face of it, this seems to be a perfect solution. I really don't care about rows that other clients are inserting/modifying, because each row belongs only to a specific client, so I may very well skip locked rows. But the MSDN documentaion makes me a bit worried - "When READPAST is specified, both row-level and page-level locks are skipped". As I said, row-level locks would not be a problem, but page-level locks may very well be, since a page might contain rows that belong to multiple clients (including the current one).
While there are lots of blog posts specifically mentioning that NOLOCK might miss rows, there seems to be none about READPAST (never) missing rows. This makes me skeptical and nervous to implement it, since there is no easy way to test it (implementing would be a piece of cake, just pop WITH(READPAST) into both statements SELECT clause and job done). Can someone confirm whether the READPAST hint can miss rows?
Workaround 3:
The final option is to use ALLOW_SNAPSHOT_ISOLATION and READ_COMMITED_SNAPSHOT. This would seem to be the only option to work 100% - at least I can't find any information that would contradict with it. But it is a little bit trickier to setup (I don't care much about the performance hit), because I'm using LINQ. Off the top of my head I probably need to manually open a SQL connection and pass it to the LINQ2SQL DataContext, etc... I haven't looked into the specifics very deeply.
Mostly I would prefer option 2 if somone could only reassure me that READPAST will never miss rows concerning the current client (as I said before, each client has and only ever deals with it's own set of rows). Otherwise I'll likely have to implement option 3, since option 1 is probably impossible...
I'll post the table definitions for the three tables as well, just in case:
CREATE TABLE [dbo].[UpdatedRows](
[Id] [uniqueidentifier] NOT NULL ROWGUIDCOL DEFAULT NEWSEQUENTIALID() PRIMARY KEY CLUSTERED,
[RowId] [uniqueidentifier] NOT NULL,
[UpdateDateTime] [datetime] NOT NULL,
) ON [PRIMARY]
GO
CREATE NONCLUSTERED INDEX IX_RowId ON dbo.UpdatedRows
([RowId] ASC) WITH (STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
GO
CREATE TABLE [dbo].[ReceivedUpdatedRows](
[Id] [uniqueidentifier] NOT NULL ROWGUIDCOL DEFAULT NEWSEQUENTIALID() PRIMARY KEY NONCLUSTERED,
[UpdatedRowId] [uniqueidentifier] NOT NULL REFERENCES [dbo].[UpdatedRows] ([Id]),
[StudioId] [uniqueidentifier] NOT NULL REFERENCES,
[ReceiveDateTime] [datetime] NOT NULL,
) ON [PRIMARY]
GO
CREATE CLUSTERED INDEX IX_Studios ON dbo.ReceivedUpdatedRows
([StudioId] ASC) WITH (STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
GO
CREATE TABLE [dbo].[SynchronizingRows](
[StudioId] [uniqueidentifier] NOT NULL
[UpdatedRowId] [uniqueidentifier] NOT NULL REFERENCES [dbo].[UpdatedRows] ([Id])
PRIMARY KEY CLUSTERED ([StudioId], [UpdatedRowId])
) ON [PRIMARY]
GO
PS! Studio = Client.
PS2! I just noticed that the index definitions have ALLOW_PAGE_LOCK=ON. If I would turn it off, would that make any difference to READPAST? Are there any negative downsides for turning it off?