Michael J. Swart

In the past I’ve written about monitoring identity columns to ensure there’s room to grow.

But there’s a related danger that’s a little more subtle. Say you have a table whose identity column is an 8-byte bigint. An application that converts those values to a 4-byte integer will not always fail! Those applications will only fail if the value is larger than 2,147,483,647.

If the conversion of a large value is done in C#, you’ll get an Overflow Exception or an Invalid Cast Exception and if the conversion is done in SQL Server you’ll see get this error message:

Msg 8115, Level 16, State 2, Line 21
Arithmetic overflow error converting expression to data type int.

The danger

If such conversions exist in your application, you won’t see any problems until the bigint identity values are larger than 2,147,483,647. My advice then is to test your application with large identity values in a test environment. But how?

Use this script to set large values on BIGINT identity columns

On a test server, run this script to get commands to adjust bigint identity values to beyond the maximum value of an integer:

-- increase bigint identity columns
select 
	'DBCC CHECKIDENT(''' + 
	QUOTENAME(OBJECT_SCHEMA_NAME(object_id)) + '.' +
	QUOTENAME(object_Name(object_id)) + ''', RESEED, 2147483648);
' as script
from 
	sys.identity_columns
where 
	system_type_id = 127
	and object_id in (select object_id from sys.tables);
 
-- increase bigint sequences
select 
	'ALTER SEQUENCE ' +
	QUOTENAME(OBJECT_SCHEMA_NAME(object_id)) + '.' +
	QUOTENAME(object_Name(object_id)) + ' 
	RESTART WITH 2147483648 INCREMENT BY ' + 
	CAST(increment as sysname) +
	' NO MINVALUE NO MAXVALUE;
' as script
from 
	sys.sequences
where 
	system_type_id = 127;

Prepared for testing

The identity columns in your test database are now prepared for testing. And hopefully you have an automated way to exercise your application code to find sneaky conversions to 4-byte integers. I found several of these hidden defects myself and I’m really glad I had the opportunity to tackle these before they became an issue in production.

When I wrote Take Care When Scripting Batches, I wanted to guard against a common pitfall when implementing a batching solution (n-squared performance). I suggested a way to be careful. But I knew that my solution was not going to be universally applicable to everyone else’s situation. So I wrote that post with a focus on how to evaluate candidate solutions.

But we developers love recipes for problem solving. I wish it was the case that for whatever kind of problem you got, you just stick the right formula in and problem solved. But unfortunately everyone’s situation is different and the majority of questions I get are of the form “What about my situation?” I’m afraid that without extra details, the best advice remains to do the work to set up the tests and find out for yourself.

But despite that. I’m still going to answer some common questions I get. But I’m going to continue to focus on how I evaluate each solution.
(Before reading further, you might want to re-familiarize yourself with the original article Take Care When Scripting Batches).

Here are some questions I get:

What if the clustered index is not unique?

Or what if the clustered index had more than one column such that leading column was not unique. For example, imagine the table was created with this clustered primary key:

ALTER TABLE dbo.FactOnlineSales
ADD CONSTRAINT PK_FactOnlineSales
PRIMARY KEY CLUSTERED (DateKey, OnlineSalesKey)

How do we write a batching script in that case? It’s usually okay if you just use the leading column of the clustered index. The careful batching script looks like this now:

DECLARE
  @LargestKeyProcessed DATETIME = '20000101',
  @NextBatchMax DATETIME,
  @RC INT = 1;
 
WHILE (@RC > 0)
BEGIN
 
  SELECT TOP (1000) @NextBatchMax = DateKey
  FROM dbo.FactOnlineSales
  WHERE DateKey > @LargestKeyProcessed
    AND CustomerKey = 19036
  ORDER BY DateKey ASC;
 
  DELETE dbo.FactOnlineSales
  WHERE CustomerKey = 19036
    AND DateKey > @LargestKeyProcessed
    AND DateKey <= @NextBatchMax;
 
  SET @RC = @@ROWCOUNT;
  SET @LargestKeyProcessed = @NextBatchMax;
 
END

The performance is definitely comparable to the original careful batching script:

Logical Reads Per Delete

But is it correct? A lot of people wonder if the non-unique index breaks the batching somehow. And the answer is yes, but it doesn’t matter too much.

By limiting the batches by DateKey instead of the unique OnlineSalesKey, we are giving up batches that are exactly 1000 rows each. In fact, most of the batches in my test process somewhere between 1000 and 1100 rows and the whole thing requires three fewer batches than the original script. That’s acceptable to me.

If I know that the leading column of the clustering key is selective enough to keep the batch sizes pretty close to the target size, then the script is still accomplishing its goal.

What if the rows I have to delete are sparse?

Here’s another situation. What if instead of customer 19036, we were asked to delete customer 7665? This time, instead of deleting 45100 rows, we only have to delete 379 rows.

I try the careful batching script and see that all rows are deleted in a single batch. SQL Server was looking for batches of 1000 rows to delete. But since there aren’t that many, it scanned the entire table to find just 379 rows. It completed in one batch, but that single batch performed as poorly as the straight algorithm.

One solution is to create an index (online!) for these rows. Something like:

CREATE INDEX IX_CustomerKey 
ON dbo.FactOnlineSales(CustomerKey) 
WITH (ONLINE = ON);

Most batching scripts are one-time use. So maybe this index is one-time use as well. If it’s a temporary index, just remember to drop it after the script is complete. A temp table could also do the same trick.

With the index, the straight query only needed 3447 logical reads to find all the rows to delete:

DELETE dbo.FactOnlineSales WHERE CustomerKey = 7665;

Logical Reads

Can I use the Naive algorithm if I use a new index?

How does the Naive and other algorithms fare with this new index on dbo.FactOnlineSales(CustomerKey)?

The rows are now so easy to find that the Naive algorithm no longer has the n-squared behavior we worried about earlier. But there is some extra overhead. We have to delete from more than one index. And we’re doing many b-tree lookups (instead of just scanning a clustered index).

Remember the Naive solution looks like this:

DECLARE	@RC INT = 1;
 
WHILE (@RC > 0)
BEGIN
 
  DELETE TOP (1000) dbo.FactOnlineSales
  WHERE CustomerKey = 19036;
 
  SET @RC = @@ROWCOUNT
 
END

But now with the index, the performance looks like this (category Naive with Index)

The index definitely helps. With the index, the Naive algorithm definitely looks better than it did without the index. But it still looks worse than the careful batching algorithm.

But look at that consistency! Each batch processes 1000 rows and reads exactly the same amount. I might choose to use Naive batching with an index if I don’t know how sparse the rows I’m deleting are. There are a lot of benefits to having a constant runtime for each batch when I can’t guarantee that rows aren’t sparse.

Explore new solutions on your own

There are many different solutions I haven’t explored. This list isn’t comprehensive.

But it’s all tradeoffs. When faced with a choice between candidate solutions, it’s essential to know how to test and measure each solution. SQL Server has more authoritative answers about the behavior of SQL Server than me or any one else. Good luck.

System procedures like sp_replincrementlsn and system functions like fn_cdc_get_min_lsn and fn_cdc_get_max_lsn return values that are of type binary(10).

These values represent LSNs, Log Sequence Numbers which are an internal way to represent the ordering of transaction logs.

Typically as developers, we don’t care about these values. But when we do want to dig into the transaction log, we can do so with sys.fn_dblog which takes two optional parameters. These parameters are LSN values which limit the results of sys.fn_dblog. But the weird thing is that sys.fn_dblogis a function whose LSN parameters are NVARCHAR(25).

The function sys.fn_dblog doesn’t expect binary(10) values for its LSN parameters, it wants the LSN values as a formatted string, something like: 0x00000029:00001a3c:0002.

Well, to convert the binary(10) LSN values into the format expected by sys.fn_dblog, I came up with this function:

CREATE OR ALTER FUNCTION dbo.fn_lsn_to_dblog_parameter(
    @lsn BINARY(10)
)
RETURNS NVARCHAR(25)
AS 
BEGIN
  RETURN
    NULLIF(
      STUFF (
        STUFF (
          '0x' + CONVERT(NVARCHAR(25), @lsn, 2),
          11, 0, ':' ),
        20, 0, ':' ),
      '0x00000000:00000000:0000'
    )
END
GO

Example

I can increment the LSN once with a no-op and get back the lsn value with sp_replincrementlsn.
I can then use fn_lsn_to_dblog_parameter to get an LSN string to use as parameters to sys.fn_dblog.
This helps me find the exact log entry in the transaction that corresponds to that no-op:

DECLARE @lsn binary(10);
DECLARE @lsn_string nvarchar(25)
exec sp_replincrementlsn @lsn OUTPUT;
SET @lsn_string = dbo.fn_lsn_to_dblog_parameter(@lsn);
 
select @lsn_string as lsn_string, [Current LSN], Operation
from sys.fn_dblog(@lsn_string, @lsn_string);

To avoid deadlocks when implementing the upsert pattern, make sure the index on the key column is unique. It’s not sufficient that all the values in that particular column happen to be unique. The index must be defined to be unique, otherwise concurrent queries can still produce deadlocks.

Say I have a table with an index on Id (which is not unique):

CREATE TABLE dbo.UpsertTest(
	Id INT NOT NULL,
	IdString VARCHAR(100) NOT NULL,
	INDEX IX_UpsertTest CLUSTERED (Id)
)

I implement my test UPSERT procedure the way I’m supposed to like this:

CREATE OR ALTER PROCEDURE dbo.s_DoSomething  
AS 
SET TRANSACTION ISOLATION LEVEL SERIALIZABLE
BEGIN TRANSACTION 
	DECLARE @Id BIGINT = DATEPART(SECOND, GETDATE());
	DECLARE @IdString VARCHAR(100) = CAST(@Id AS VARCHAR(100)); 
 
	IF EXISTS ( 
		SELECT * 
		FROM dbo.UpsertTest WITH (UPDLOCK) 
		WHERE Id = @Id 
	) 
	BEGIN 
		UPDATE dbo.UpsertTest 
		SET IdString = @IdString 
		WHERE Id = @Id; 
	END 
	ELSE 
	BEGIN 
		INSERT dbo.UpsertTest (Id, IdString) 
		VALUES (@Id, @IdString); 
	END; 
COMMIT

When I exercise this procedure concurrently with many threads it produces deadlocks! I can use extended events and the output from trace flag 1200 to find out what locks are taken and what order.

What Locks Are Taken?

It depends on the result of the IF statement. There are two main scenarios to look at. Either the row exists or it doesn’t.

Scenario A: The Row Does Not Exist (Insert)
These are the locks that are taken:

For the IF EXISTS statement:
• Acquire Range S-U lock on resource (ffffffffffff) which represents “infinity”

For the Insert statement:

• Acquire RangeI-N lock on resource (ffffffffffff)
• Acquire X lock on resource (66467284bfa8) which represents the newly inserted row

Scenario B: The Row Exists (Update)
The locks that are taken are:

For the IF EXISTS statement:
• Acquire Range S-U lock on resource (66467284bfa8)

For the Update statement:

• Acquire RangeX-X lock on resource (66467284bfa8)
• Acquire RangeX-X lock on resource (ffffffffffff)

Scenario C: The Row Does Not Exist, But Another Process Inserts First (Update)
There’s a bonus scenario that begins just like the Insert scenario, but the process is blocked waiting for resource (ffffffffffff). Once it finally acquires the lock, the next locks that are taken look the same as the other Update scenario. The locks that are taken are:

For the IF EXISTS statement:
• Wait for Range S-U lock on resource (ffffffffffff)
• Acquire Range S-U lock on resource (ffffffffffff)
• Acquire Range S-U lock on resource (66467284bfa8)

For the Update statement:

• Acquire RangeX-X lock on resource (66467284bfa8)
• Acquire RangeX-X lock on resource (ffffffffffff)

The Deadlock

And when I look at the deadlock graph, I can see that the two update scenarios (Scenario B and C) are fighting:
Scenario B:

• Acquire RangeX-X lock on resource (66467284bfa8) during UPDATE
• Blocked RangeX-X lock on resource (ffffffffffff) during UPDATE

Scenario C:

• Acquire RangeS-U lock on resource (ffffffffffff) during IF EXISTS
• Blocked RangeS-U lock on resource (66467284bfa8) during IF EXISTS

Why Isn’t This A Problem With Unique Indexes?

To find out, let’s take a look at one last scenario where the index is unique:
Scenario D: The Row Exists (Update on Unique Index)

For the IF EXISTS statement:
• Acquire U lock on resource (66467284bfa8)

For the Update statement:

• Acquire X lock on resource (66467284bfa8)

Visually, I can compare scenario B with Scenario D:

When the index is not unique, SQL Server has to take key-range locks on either side of the row to prevent phantom inserts, but it’s not necessary when the values are guaranteed to be unique! And that makes all the difference. When the index is unique, no lock is required on resource (ffffffffffff). There is no longer any potential for a deadlock.

Solution: Define Indexes As Unique When Possible

Even if the values in a column are unique in practice, you’ll help improve concurrency by defining the index as unique. This tip can be generalized to other deadlocks. Next time you’re troubleshooting a deadlock involving range locks, check to see whether the participating indexes are unique.

This quirk of requiring unique indexes for the UPSERT pattern is not unique to SQL Server, I notice that PostgreSQL requires a unique index when using their “ON CONFLICT … UPDATE” syntax. This is something they chose to do very deliberately.

Other Things I Tried

This post actually comes from a real problem I was presented. It took a while to reproduce and I tried a few things before I settled on making my index unique.

Lock More During IF EXISTS?
Notice that there is only one range lock taken during the IF EXISTS statement, but there are two range needed for the UPDATE statement. Why is only one needed for the EXISTS statement? If extra rows get inserted above the row that was read, that doesn’t change the answer to EXISTS. So it’s technically not a phantom read and so SQL Server doesn’t take that lock.

So what if I changed my IF EXISTS to

IF ( 
	SELECT COUNT(*)
	FROM dbo.UpsertTest WITH (UPDLOCK) 
	WHERE Id = @Id 
) > 0

That IF statement now takes two range locks which is good, but it still gets tripped up with Scenario C and continues to deadlock.

Update Less?
Change the update statement to only update one row using TOP (1)

UPDATE TOP (1) dbo.UpsertTest 
SET IdString = @IdString
WHERE Id = @Id;

During the update statement, this only requires one RangeX-X lock instead of two. And that technique actually works! I was unable to reproduce deadlocks with TOP (1). So it is indeed a candidate solution, but making the index unique is still my preferred method.

The configuration setting cost threshold for parallelism has a default value of 5. As a default value, it’s probably too low and should be raised. But what benefit are we hoping for? And how can we measure it?

The database that I work with is a busy OLTP system with lots of very frequent, very inexpensive queries and so I don’t like to see any query that needs to go parallel.

What I’d like to do is raise the configuration cost threshold to something larger and look at the queries that have gone from multi-threaded to single-threaded. I want to see that these queries become cheaper on average. By cheaper I mean consume less cpu. I expect the average duration of these queries to increase.

How do I find these queries? I can look in the cache. The view sys.dm_exec_query_stats can tell me if a query plan is parallel, and I can look into the plans for the estimated cost. In my case, I have relatively few parallel queries. Only about 300 which means the xml parsing piece of this query runs pretty quickly.

Measure the Cost of Parallel Queries

WITH XMLNAMESPACES (DEFAULT 'http://schemas.microsoft.com/sqlserver/2004/07/showplan')
SELECT 
	sql_text.[text] as sqltext,
	qp.query_plan,
	xml_values.subtree_cost as estimated_query_cost_in_query_bucks,
	qs.last_dop,
	CAST( qs.total_worker_time / (qs.execution_count + 0.0) as money ) as average_query_cpu_in_microseconds,
	qs.total_worker_time,
	qs.execution_count,
	qs.query_hash,
	qs.query_plan_hash,
	qs.plan_handle,
	qs.sql_handle	
FROM sys.dm_exec_query_stats qs
CROSS APPLY sys.dm_exec_sql_text(qs.sql_handle) st
CROSS APPLY sys.dm_exec_query_plan (qs.plan_handle) qp
CROSS APPLY 
	(
		SELECT SUBSTRING(st.[text],(qs.statement_start_offset + 2) / 2,
		(CASE 
			WHEN qs.statement_end_offset = -1  THEN LEN(CONVERT(NVARCHAR(MAX),st.[text])) * 2
			ELSE qs.statement_end_offset + 2
			END - qs.statement_start_offset) / 2)
	) as sql_text([text])
OUTER APPLY 
	( 
		SELECT 
			n.c.value('@QueryHash', 'NVARCHAR(30)')  as query_hash,
			n.c.value('@StatementSubTreeCost', 'FLOAT')  as subtree_cost
		FROM qp.query_plan.nodes('//StmtSimple') as n(c)
	) xml_values
WHERE qs.last_dop > 1
AND sys.fn_varbintohexstr(qs.query_hash) = xml_values.query_hash
AND execution_count > 10
ORDER BY xml_values.subtree_cost
OPTION (RECOMPILE);

What Next?

Keep track of the queries you see whose estimated subtree cost is below the new threshold you’re considering. Especially keep track of the query_hash and the average_query_cpu_in_microseconds.
Then make the change and compare the average_query_cpu_in_microseconds before and after. Remember to use the sql_hash as the key because the plan_hash will have changed.
Here’s the query modified to return the “after” results:

Measure the Cost of Those Queries After Config Change

WITH XMLNAMESPACES (DEFAULT 'http://schemas.microsoft.com/sqlserver/2004/07/showplan')
SELECT 
	sql_text.[text] as sqltext,
	qp.query_plan,
	xml_values.subtree_cost as estimated_query_cost_in_query_bucks,
	qs.last_dop,
	CAST( qs.total_worker_time / (qs.execution_count + 0.0) as money ) as average_query_cpu_in_microseconds,
	qs.total_worker_time,
	qs.execution_count,
	qs.query_hash,
	qs.query_plan_hash,
	qs.plan_handle,
	qs.sql_handle	
FROM sys.dm_exec_query_stats qs
CROSS APPLY sys.dm_exec_sql_text(qs.sql_handle) st
CROSS APPLY sys.dm_exec_query_plan (qs.plan_handle) qp
CROSS APPLY 
	(
		SELECT SUBSTRING(st.[text],(qs.statement_start_offset + 2) / 2,
		(CASE 
			WHEN qs.statement_end_offset = -1  THEN LEN(CONVERT(NVARCHAR(MAX),st.[text])) * 2
			ELSE qs.statement_end_offset + 2
			END - qs.statement_start_offset) / 2)
	) as sql_text([text])
OUTER APPLY 
	( 
		SELECT 
			n.c.value('@QueryHash', 'NVARCHAR(30)')  as query_hash,
			n.c.value('@StatementSubTreeCost', 'FLOAT')  as subtree_cost
		FROM qp.query_plan.nodes('//StmtSimple') as n(c)
	) xml_values
WHERE qs.query_hash in ( /* put the list of sql_handles you saw from before the config change here */ )
AND sys.fn_varbintohexstr(qs.query_hash) = xml_values.query_hash
ORDER BY xml_values.subtree_cost
OPTION (RECOMPILE);

What I Found

In general, increasing the threshold from 5 –> 50 generally had a very good effect on those queries that went from multithreaded to singlethreaded.
Over half of the queries improved at least 1 order of magnitude (and a couple improved 3 orders of magnitude!) https://t.co/iELhiSr7On pic.twitter.com/W4Mv7hGVlz

— Michael J Swart (@MJSwart) January 19, 2022