Michael J. Swart

August 9, 2022

Formatting Binary(10) LSN Values For Use In sys.fn_dblog()

Filed under: Miscelleaneous SQL,SQL Scripts,Technical Articles — Michael J. Swart @ 3:06 pm

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);

March 18, 2022

UPSERT Requires a Unique Index

Filed under: Miscelleaneous SQL,SQL Scripts,Technical Articles — Michael J. Swart @ 10:11 am

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 Update statement:

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

Insert Scenario

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)

Update Scenario

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)

Update After Wait Scenario

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:
Update Two Scenarios

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.

February 7, 2022

Five Ways Time Makes Unit Tests Flaky

Filed under: Miscelleaneous SQL,Technical Articles — Michael J. Swart @ 12:21 pm
Five Ways Time Makes Unit Tests Flaky
I explore different sources of test flakiness related to time:

A flaky test is a unit test that sometimes passes and sometimes fails. The causes of these flaky tests are often elusive because they’re not consistently reproducible.

I’ve found that unit tests that deal with dates and times are notorious for being flaky – especially such tests that talk to SQL Server. I want to explore some of the reasons this can happen.

My Setup


All scripts and code samples are available on github.
In the examples I discuss below, I’m using a table defined like this:

CREATE TABLE dbo.MESSAGE_LOG
(
	LogId INT IDENTITY NOT NULL 
		PRIMARY KEY,
	LogMessage NVARCHAR(MAX) NOT NULL,
	CreatedDate DATETIME
		DEFAULT (SYSDATETIME())
)

I also wrote some methods that execute these queries:

AddLogMessage

INSERT dbo.MESSAGE_LOG(LogMessage)
OUTPUT inserted.LogId
VALUES (@Message);

AddLogMessageWithDate
Same method but this allows the application to supply the LastUpdate value

INSERT dbo.MESSAGE_LOG(LogMessage, LastUpdate)
OUTPUT inserted.LogId
VALUES (@Message, @LastUpdate)

UpdateLogMessage

UPDATE dbo.MESSAGE_LOG
SET Message = @Message,
    LastUpdatedTime = SYSDATETIME()
WHERE LogId = @LogId

Sources of Flaky Tests

In no particular order:

Tests Run Too Quick?

The following test checks to see that UpdateMessage updated the LastUpdate column.

[Test]
public void UpdateMessage_DateIsUpdated_1() {
    string message = Guid.NewGuid().ToString();
    int logId = m_data.AddLogMessage( message );
    LogMessageDto? dto = m_data.GetLogMessage( logId );
    DateTime createdDate = dto.LastUpdate;
 
    string newMessage = Guid.NewGuid().ToString();
    m_data.UpdateLogMessage( logId, newMessage );
 
    dto = m_data.GetLogMessage( logId );
    DateTime updatedDate = dto.LastUpdate;
 
    // The following assertion may fail! 
    // updatedDate and createdDate are Equal if the server is fast enough
    Assert.Greater( updatedDate, createdDate ); 
}

The test ran so quickly that updatedDate has the same value as createdDate. This test may fail with this error message:

    Failed UpdateMessage_DateIsUpdated [55 ms]
    Error Message:
    Expected: greater than 2022-02-05 15:18:10.33
    But was: 2022-02-05 15:18:10.33

It’s tempting to to get around this by adding a Thread.Sleep call between the insert and update. I don’t recommend it. That kind of pattern adds up and really lengthens the time it takes to run all tests.

Another solution might involve changing Greater to GreaterOrEqual but then we can’t verify that the value has actually been updated.

Storing dates using a more precise datetime type like DATETIME2 may help avoid more failures, but maybe not all failures.

The Right Way
Ideally we want to set up the test case such that the LastUpdate value is a constant date that’s definitely in the past. I would change this test to use AddLogMessageWithDate instead of AddLogMessage:

    DateTime then = new DateTime(2000, 01, 01);
    int logId = m_data.AddLogMessageWithDate( message, then );

Not All DateTimes Are Created Equal


.Net’s DateTime is different than SQL Server’s DATETIME. Specifically they have different precisions. DateTime values in SQL Server are rounded to increments of .000, .003, or .007 seconds. This means that you can’t store a .Net DateTime value in SQL Server and get it back. This test demonstrates the problem:

[Test]    
public void StoreDate_ReadItBack() {
    // Store date
    string message = Guid.NewGuid().ToString();
    DateTime now = DateTime.Now;
    int logId = m_data.AddLogMessageWithDate( message, now );
 
    // Read it back
    LogMessageDto? dto = m_data.GetLogMessage( logId );
 
    // The following assertion may fail! 
    // SQL Server's DATETIME has a different precision than .Net's DateTime
    Assert.AreEqual( now, dto.LastUpdate );
}

It may fail with:

    Failed StoreDate_ReadItBack [101 ms]
    Error Message:
    Expected: 2022-02-04 15:11:20.4474577
    But was: 2022-02-04 15:11:20.447

The Right Way
Understanding the resolution limitations of SQL Server’s DateTime is important here. A few solutions come to mind:

  • Maybe use a constant value instead of “now”
  • Modify the database columns to use SQL Server’s DATETIME2 which has a better resolution
  • Just fetch “now” from the database. I like this idea. When I use it again later, I’ll go into more detail.

Time Zones (Of Course)


Running integration tests that talk to a database on a separate servercan mean translating server times back and forth between both servers. This leads to another common source of flakiness: time zones. It’s not easy to avoid this kind of issue. Both Azure and AWS try to tackle this by use UTC everywhere.

A flaky test might look like this.

public void UpdateMessage_DateIsUpdated_2() {
    string message = Guid.NewGuid().ToString();
    DateTime now = DateTime.Now;
    int logId = m_data.AddLogMessageWithDate( message, now );
 
    string newMessage = Guid.NewGuid().ToString();
    m_data.UpdateLogMessage( logId, newMessage );
 
    LogMessageDto? dto = m_data.GetLogMessage( logId );
 
    // This next assertion can fail if the database is in a different time zone        
    Assert.GreaterOrEqual( dto.LastUpdate, now );
}

It fails like this:

    Failed UpdateMessage_DateIsUpdated_2 [19 ms]
    Error Message:
    Expected: greater than or equal to 2022-02-05 21:06:54.521464
    But was: 2022-02-05 16:06:54.52

Why is this pattern a source of flaky tests? The success of the test depends on the time zones of the test server and the database server. But even if you control both time zones, this particular example is still vulnerable to clock drift as we’ll see later.

The Right Way
Use a constant time or try fetching “now” from the database.

DateTime now = m_nowProvider.Now();

Here I’m using a method I wrote which simply returns the value of SELECT GETDATE(); from the database.

Clock Drift


Related to time zones is clock drift which again causes errors when you compare dates from two different servers.

No server’s clock is perfect and I like to think of each server’s clock as having its own time zone. Windows tells me that my laptop is set at (UTC -05:00) but with clock drift it’s probably something like (UTC -05:00:01.3). You can work at synchronizing clocks, but unless you’re testing that synchronization, you shouldn’t depend on it in your tests.

Just like in the case with time zones, this test may fail when it compares times from two different clocks:

public void UpdateMessage_DateIsUpdated_3() {
    string message = Guid.NewGuid().ToString();
    DateTime now = DateTime.Now;
    int logId = m_data.AddLogMessageWithDate( message, now );
 
    string newMessage = Guid.NewGuid().ToString();
    m_data.UpdateLogMessage( logId, newMessage );
 
    LogMessageDto? dto = m_data.GetLogMessage( logId );
 
    // This next test can fail if the clocks on the database server is off by a few seconds
    Assert.GreaterOrEqual( dto.LastUpdate, now );
}

The Right Way
Just like before, use a constant value or try fetching “now” from the database.

DateTime now = m_nowProvider.Now();

This way we’re comparing times from only one server’s clock.

Daylight Savings (Of Course)


This next test is flaky because of daylight savings time. It’s not specific to SQL Server but I thought I’d include it because I have been burned by this before:

[Test]    
public void StoreDateInTheFuture() {
    string message = Guid.NewGuid().ToString();
    DateTime inAMonth = DateTime.Now + TimeSpan.FromDays( 30 );        
 
    // CovertTime may fail because "a month from now" may be an invalid DateTime (with daylight savings)
    inAMonth = TimeZoneInfo.ConvertTime( inAMonth, TimeZoneInfo.Local );
    m_data.AddLogMessageWithDate( message, inAMonth );
    Assert.Pass();
}

I saw a test just like this one fail at 2:18 AM on February 9th, 2018. Adding 30 days to that date brought us to 2:18AM which was right in the middle of the hour we were skipping for daylight savings time and that’s what caused the error. This test fails with:

    Failed StoreDateInTheFuture [32 ms]
    Error Message:
    System.ArgumentException : The supplied DateTime represents an invalid time. For example, when the clock is adjusted forward, any time in the period that is skipped is invalid. (Parameter ‘dateTime’)

Summary


Flaky tests come from non-deterministic tests. To quote Martin Fowler, “Few things are more non-deterministic than a call to the system clock”. Try to:

  • Write tests with hard coded dates
  • Avoid comparing dates sourced from two different clocks
  • Consider writing a “NowProvider” (which can be mocked!)
  • Be very deliberate about time zones
  • Be very deliberate about data types (both in C# and SQL Server)

January 19, 2022

Measure the Effect of “Cost Threshold for Parallelism”

Filed under: Miscelleaneous SQL,SQL Scripts,Technical Articles — Michael J. Swart @ 10:40 am

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

August 9, 2021

Find Procedures That Use SELECT *

Filed under: Miscelleaneous SQL,SQL Scripts,SQLServerPedia Syndication,Technical Articles — Michael J. Swart @ 12:00 pm

I have trouble with procedures that use SELECT *. They are often not “Blue-Green safe“. In other words, if a procedure has a query that uses SELECT * then I can’t change the underlying tables can’t change without causing some tricky deployment issues. (The same is not true for ad hoc queries from the application).

I also have a lot of procedures to look at (about 5000) and I’d like to find the procedures that use SELECT *.
I want to maybe ignore SELECT * when selecting from a subquery with a well-defined column list.
I also want to maybe include related queries like OUTPUT inserted.*.

The Plan

  1. So I’m going to make a schema-only copy of the database to work with.
  2. I’m going to add a new dummy-column to every single table.
  3. I’m going to use sys.dm_exec_describe_first_result_set_for_object to look for any of the new columns I created

Any of my new columns that show up, were selected with SELECT *.

The Script

use master;
DROP DATABASE IF EXISTS search_for_select_star;
DBCC CLONEDATABASE (the_name_of_the_database_you_want_to_analyze, search_for_select_star);
ALTER DATABASE search_for_select_star SET READ_WRITE;
GO
 
use search_for_select_star;
 
DECLARE @SQL NVARCHAR(MAX);
SELECT 
	@SQL = STRING_AGG(
		CAST(
			'ALTER TABLE ' + 
			QUOTENAME(OBJECT_SCHEMA_NAME(object_id)) + 
			'.' + 
			QUOTENAME(OBJECT_NAME(object_id)) + 
			' ADD NewDummyColumn BIT NULL' AS NVARCHAR(MAX)),
		N';')
FROM 
	sys.tables;
 
exec sp_executesql @SQL;
 
SELECT 
	SCHEMA_NAME(p.schema_id) + '.' + p.name AS procedure_name, 
	r.column_ordinal,
	r.name
FROM 
	sys.procedures p
CROSS APPLY 
	sys.dm_exec_describe_first_result_set_for_object(p.object_id, NULL) r
WHERE 
	r.name = 'NewDummyColumn'
ORDER BY 
	p.schema_id, p.name;
 
use master;
DROP DATABASE IF EXISTS search_for_select_star;

Update

Tom from StraightforwardSQL pointed out a nifty feature that Microsoft has already implemented.

Yes it does! You can use it like this:

select distinct SCHEMA_NAME(p.schema_id) + '.' + p.name AS procedure_name
from sys.procedures p
cross apply sys.dm_sql_referenced_entities(
	object_schema_name(object_id) + '.' + object_name(object_id), default) re
where re.is_select_all = 1

Comparing the two, I noticed that my query – the one that uses dm_exec_describe_first_result_set_for_object – has some drawbacks. Maybe the SELECT * isn’t actually included in the first result set, but some subsequent result set. Or maybe the result set couldn’t be described for one of these various reasons

On the other hand, I noticed that dm_sql_referenced_entities has a couple drawbacks itself. It doesn’t seem to capture select statements that use `OUTPUT INSERTED.*` for example.

In practice though, I found the query that Tom suggested works a bit better. In the product I work most closely with, dm_sql_referenced_entities only missed 3 procedures that dm_exec_describe_first_result_set_for_object caught. But dm_exec_describe_first_result_set_for_object missed 49 procedures that dm_sql_referenced_entities caught!

April 1, 2021

Only UPDATE Rows That Are Changing, But Do It Carefully

Filed under: Miscelleaneous SQL,Technical Articles,Tongue In Cheek — Michael J. Swart @ 12:20 pm

If you update a column to the exact same value as it had before, there’s still work being done.

Quite obediently, SQL Server takes out its eraser, erases the old value, and writes the same value in its place even though nothing changed!

But it feels like a real change. It has consequences for locking and an impact to the transaction log just as if it were a real change.

So that leads to performance optimizations that look like this:

Original Update Statement:

UPDATE Users
SET    DisplayName = @NewDisplayName
WHERE  Id = @Id;

Only Update When Necessary:

UPDATE Users
SET    DisplayName = @NewDisplayName
WHERE  Id = @Id
AND    DisplayName <> @NewDisplayName;

But Take Care!

Be careful of this kind of optimization. For example, you have to double check that DisplayName is not a nullable column (do you know why?). There are other things to worry about too, mostly side effects:

Side Effects

This simple update statement can have loads of side effects that can be hard to see. And the trouble with any side effect, is that other people can place dependencies on them! It happens all the time. Here is a list of just some of the side effects I can think of, I’m sure it’s not exhaustive.

Triggers: Ugh, I dislike triggers at the best of times, so check out any triggers that might exist on the table. In the original UPDATE statement, the row always appears in the INSERTED and DELETED tables, but in the improved version, the row does not necessarily. You have to see if that matters.

RowCount: What if the original update statement was part of a batch that looked like this:

UPDATE Users
SET    DisplayName = @NewDisplayName
WHERE  Id = @Id;
 
IF (@@ROWCOUNT > 0)
    RAISERROR ('Could not find User to update', 16, 1);

At least this side effect has the benefit of not being hidden. It’s located right beside the code that it depends on.

Rowversion: A rowversion value changes every time a row changes. Such a column would get updated in the original UPDATE statement, but not in the improved version. I can think of a number of reasonable of use cases that might depend on a rowversion column. ETLs for example that only care about changed data. So this might actually be an improvement for that ETL, but then again, maybe the number of “changed” rows was the important part and that number is now changing with the improvement. Speaking of ETLs:

Temporal Tables: Yep, the UPDATE statement is a “change” in the table that gets tracked in temporal history.

Change Data Capture, etc…: I haven’t bothered to set up Change Data Capture to check, but I assume that an UPDATE statement that updates a row to the same value is still considered a change. Right or wrong, the performance improvement changes that assumption.

People Depend On Side Effects

When I see people do this, I start to feel grouchy: Someone’s getting in the way of my performance improvement! But it happens. People depend on side effects like these all the time. I’m sure I do. XKCD pokes fun at this with Workflow where he notices that “Every change breaks someone’s workflow”. And now I’m imagining a case where some knucklehead is using the growth of the transaction log as a metric, like “Wow, business is really booming today, 5GB of transaction log growth and it’s not even noon!”

Although these are silly examples, there are of course more legit examples I could probably think of. And so in a well-functioning organization, we can’t unilaterally bust other peoples workflows (as much as we might like to).

January 26, 2021

Avoid This Pitfall When Using sp_getapplock

Filed under: Miscelleaneous SQL,SQL Scripts,SQLServerPedia Syndication,Technical Articles — Michael J. Swart @ 11:10 am

Takeaway: For most use cases, using sp_releaseapplock is unnecessary. Especially when using @LockOwner = 'Transaction (which is the default).

The procedure sp_getapplock is a system stored procedure that can be helpful when developing SQL for concurrency. It takes a lock on an imaginary resource and it can be used to avoid race conditions.

But I don’t use sp_getapplock a lot. I almost always depend on SQL Server’s normal locking of resources (like tables, indexes, rows etc…). But I might consider it for complicated situations (like managing sort order in a hierarchy using a table with many different indexes).

In that case, I might use it something like this:

BEGIN TRAN
 
exec sp_getapplock 
    @Resource = @LockResourceName, 
    @LockMode = 'Exclusive', 
    @LockOwner = 'Transaction';
 
/* read stuff (e.g. "is time slot available?") */
/* change stuff (e.g. "make reservation") */
 
exec sp_releaseapplock
    @Resource = @LockResourceName, 
    @LockOwner = 'Transaction';
 
COMMIT

But there’s a problem with this pattern, especially when using RCSI. After sp_releaseapplock is called, but before the COMMIT completes, another process running the same code can read the previous state. In the example above, both processes will think a time slot is available and will try to make the same reservation.

What I really want is to release the applock after the commit. But because I specified the lock owner is 'Transaction'. That gets done automatically when the transaction ends! So really what I want is this:

BEGIN TRAN
 
exec sp_getapplock 
    @Resource = @LockResourceName, 
    @LockMode = 'Exclusive', 
    @LockOwner = 'Transaction';
 
/* read stuff (e.g. "is time slot available?") */
/* change stuff (e.g. "make reservation") */
 
COMMIT -- all locks are freed after this commit

January 8, 2021

Collect Wait Stats Regularly in Order To Report On Them Over Time

Filed under: Miscelleaneous SQL,SQL Scripts,SQLServerPedia Syndication,Technical Articles — Michael J. Swart @ 12:48 pm

I find wait stats so useful that I’ve got Paul Randal’s SQL Server Wait Statistics (or please tell me where it hurts…) bookmarked and I visit it frequently.

But that gives the total amount of waits for each wait type accumulated since the server was started. And that isn’t ideal when I’m troubleshooting trouble that started recently. No worries, Paul also has another fantastic post Capturing wait statistics for a period of time.

Taking that idea further, I can collect data all the time and look at it historically, or just for a baseline. Lot’s of monitoring tools do this already, but here’s what I’ve written:

Mostly I’m creating these scripts for me. I’ve created a version of these a few times now and some reason, I can’t find them each time I need them again!

This stuff can be super useful, especially, if you combine it with a visualization tool (like PowerBI or even Excel).
For example, here’s a chart I made when we were experiencing the XVB_LIST spinlock issues I wrote about not too long ago. Good visualizations can really tell powerful stories.

A visualization of a spinlock graph

I’m talking here about spins and not waits of course, but the idea is the same and I’ve included the spinlock monitoring scripts in the same repo on github.

Also a quick reminder wait stats aren’t everything. Don’t neglect monitoring resources as Greg Gonzales pointed out last year.

October 28, 2020

Detect Excessive Spinlock Contention on SQL Server

Scaling SQL Server High
The beginning of the school year is behind us and what a semester start! 2020 has been tough on many of us and I’m fortunate to work for a company whose services are in such high demand. In fact we’ve seen some scaling challenges like we’ve never seen before. I want to talk about some of them.

Context

As we prepared to face unprecedented demand this year, we began to think about whether bigger is better. Worried about CPU limits, we looked to what AWS had to offer in terms of their instance sizes.

We were already running our largest SQL Servers on r5 instances with 96 logical CPUs. But we decided to evaluate the pricy u instances which have 448 logical CPUs and a huge amount of memory.

Painful Symptoms

Well, bigger is not always better. We discovered that as we increased the load on the u-series servers, there would come a point where all processors would jump to 100% and stayed there. You could say it plateaued (based on the graph, would that be a plateau? A mesa? Or a butte?)

Graph showing cpu plataued at 100%

When that occurred, the number of batch requests that the server could handle dropped significantly. So we saw more CPU use, but less work was getting done.

The high demand kept the CPU at 100% with no relief until the demand decreased. When that happened, the database seemed to recover. Throughput was restored and the database’s metrics became healthy again. During this trouble we looked at everything including the number of spins reported in the sys.dm_os_spinlock_stats dmv.

The spins and backoffs reported seemed extremely high, especially for the category “XVB_LIST”, but we didn’t really have a baseline to tell whether those numbers were problematic. Even after capturing the numbers and visualizing them we saw larger than linear increases as demand increased, but were those increases excessive?

How To Tell For Sure

Chris Adkin has a post Diagnosing Spinlock Problems By Doing The Math. He explains why spinlocks are useful. It doesn’t seem like a while loop that chews up CPU could improve performance, but it actually does when it helps avoid context switches. He gives a formula to help find how much of the total CPU is spent spinning. That percentage can then help decide whether the spinning is excessive.

But I made a tiny tweak to his formula and I wrote a script to have SQL Server do the math:

  • You still have to give the number of CPUs on your server. If you don’t have those numbers handy, you can get them from SQL Server’s log. I include one of Glenn Berry’s diagnostic queries for that.
  • There’s an assumption in Chris’s calculation that one spin consumes one CPU clock cycle. A spin is really cheap (because it can use the test-and-set instruction), but it probably consumes more than one clock cycle. I assume four, but I have no idea what the actual value is.
EXEC sys.xp_readerrorlog 0, 1, N'detected', N'socket';
-- SQL Server detected 2 sockets with 24 cores per socket ...
 
declare @Sockets int = 2;
declare @PhysicalCoresPerSocket int = 24;
declare @TicksPerSpin int = 4;
 
declare @SpinlockSnapshot TABLE ( 
    SpinLockName VARCHAR(100), 
    SpinTotal BIGINT
);
 
INSERT @SpinlockSnapshot ( SpinLockName, SpinTotal )
SELECT name, spins
FROM   sys.dm_os_spinlock_stats
WHERE  spins > 0;
 
DECLARE @Ticks bigint
SELECT @Ticks = cpu_ticks 
FROM sys.dm_os_sys_info
 
WAITFOR DELAY '00:00:10'
 
DECLARE @TotalTicksInInterval BIGINT
DECLARE @CPU_GHz NUMERIC(20, 2);
 
SELECT @TotalTicksInInterval = (cpu_ticks - @Ticks) * @Sockets * @PhysicalCoresPerSocket,
       @CPU_GHz = ( cpu_ticks - @Ticks ) / 10000000000.0
FROM sys.dm_os_sys_info;
 
SELECT ISNULL(Snap.SpinLockName, 'Total') as [Spinlock Name], 
       SUM(Stat.spins - Snap.SpinTotal) as [Spins In Interval],
       @TotalTicksInInterval as [Ticks In Interval],
       @CPU_Ghz as [Measured CPU GHz],
       100.0 * SUM(Stat.spins - Snap.SpinTotal) * @TicksPerSpin / @TotalTicksInInterval as [%]
FROM @SpinlockSnapshot Snap
JOIN sys.dm_os_spinlock_stats Stat
     ON Snap.SpinLockName = Stat.name
GROUP BY ROLLUP (Snap.SpinLockName)
HAVING SUM(Stat.spins - Snap.SpinTotal) > 0
ORDER BY [Spins In Interval] DESC;

This is what I see on a very healthy server (r5.24xlarge). The server was using 14% cpu. And .03% of that is spent spinning (or somewhere in that ballpark).

A screen shot showing an example of results

More Troubleshooting Steps

So what’s going on? What is that XVB_LIST category? Microsoft says “internal use only” But I can guess. Paul Randal talks about the related latch class Versioning Transaction List. It’s an instance-wide list that is used in the implementation of features like Read Committed Snapshot Isolation (RCSI) which we do use.

Microsoft also has a whitepaper on troubleshooting this stuff Diagnose and resolve spinlock contention on SQL Server. They actually give a technique to collect call stacks during spinlock contention in order to try and maybe glean some information about what else is going on. We did that, but we didn’t learn too much. We learned that we use RCSI with lots of concurrent queries. Something we really can’t give up on.

So Then What?

What We Did

Well, we moved away from the u instance with its hundreds of CPUs and we went back to our r5 instance with only (only!) 96 logical CPUs. We’re dealing with the limits imposed by that hardware and accepting that we can’t scale higher using that box. We’re continuing to do our darnedest to move data and activity out of SQL Server and into other solutions like DynamoDb. We’re also trying to partition our databases into different deployments which spreads the load out, but introduces a lot of other challenges.

Basically, we gave up trying to scale higher. If we did want to pursue this further (which we don’t), we’d probably contact Microsoft support to try and address this spinlock contention. We know that these conditions are sufficient (if not necessary) to see the contention we saw:

  • SQL Server 2016 SP2
  • U-series instance from Amazon
  • Highly concurrent and frequent queries (>200K batch requests per second with a good mix of writes and reads on the same tables)
  • RCSI enabled.

Thank you Erin Stellato

We reached out to Erin Stellato to help us through this issue. We did this sometime around the “Painful Symptoms” section above. We had a stressful time troubleshooting all this stuff and I really appreciate Erin guiding us through it. We learned so much.

October 23, 2020

In Memory OLTP Defeated Our Tempdb Problems

Filed under: Miscelleaneous SQL,Technical Articles — Michael J. Swart @ 10:29 am
Scaling SQL Server High
The beginning of the school year is behind us and what a semester start! 2020 has been tough on many of us and I’m fortunate to work for a company whose services are in such high demand. In fact we’ve seen some scaling challenges like we’ve never seen before. I want to talk about some of them.

At D2L, we’re the perfect candidate customer for In Memory OLTP features, but we’ve held off adopting those features for years. Our servers handle tons of super quick but super frequent queries and so we find ourselves trying to address the same scaling challenges we read about in Microsoft’s customer case studies.

But there’s only one In Memory feature in particular that I care about. It’s the Memory Optimized Table Types. Specifically, I’ve always wanted to use that feature to avoid tempdb object allocation contention. Recently I finally got my chance with a lot of success. So even though I could say I’m happy with In Memory features, I think it’s more accurate to say that I feel relieved at having finally squashed my tempdb issues.

Summary of Article

The Trouble With Tempdb

We use table valued parameters with our procedures a lot (like thousands a second). We’re lucky that the table variables are not created on each execution, they’re cached. We rely heavily on the reduced overhead that this gives us. It’s for that reason we much prefer table variables over temp tables.

But when we crank up the demand, we can still run into catastrophic trouble. When tempdb contention hits us, throughput doesn’t just plateau, it drops hard. This kind of contention we see is like a kind of traffic jam where anyone who needs to use tempdb (i.e. everyone) has to wait for it. These tempdb traffic jams are rough. We even created a lighter version of sp_whoisactive that avoids tempdb issues for times like those.

I won’t go on too long about our troubles (I’ve written about tempdb issues a few times already: 1, 2, 3, 4, 5, 6, 7). The usual advice is to increase the number of tempdb data files. We were using 48 data files and really looking hard for other options.

SQL Server 2019 has some promising options. In TEMPDB – Files and Trace Flags and Updates, Oh My! Pam Lahoud points out how SQL Server can use all the PFS pages in the tempdb data files, not just the first available one. But we couldn’t move to 2019 that quickly. So we looked at Memory Optimized Table Types to help us.

Memory Optimized Table Types Can Help

Improving temp table and table variable performance using memory optimization tells us how. Our main goal is to avoid tempdb contention and memory optimized table variables don’t use tempdb at all. As long as we can be sure that the number of rows stored in these table variables is small, it’s all pros and no cons.

But it wasn’t easy for us to implement. In 2017, I wrote about Postponing Our Use Of In Memory OLTP. There were just some challenges that we couldn’t overcome. We’re not alone in struggling with the limitations of In Memory features. But our challenges weren’t the usual limitations that folks talk about and so they’re worth exploring.

The Challenge of Sardines and Whales

We have one product that we deploy to many clients. Each client gets their own database. The big ones (whales) have their own servers but the small ones (sardines) get grouped together.

Sardines and Whales

So the overhead of enabling In Memory on all the sardines was going to cause issues. The In Memory OLTP filegroup requires up to 4GB of disk space which isn’t easy to handle with hundreds of sardines. So we’re left with this dilemma. We’d like to use In Memory on the biggest whales, but not on the sardines. We tackled that in two ways

  1. Decide that the choice to add the In Memory filegroup is configuration, not product. This still required some changes though. Our backup and restore processes needed to at least handle the new filegroups, but they couldn’t expect it.
  2. Add an exception to our processes that allow schema drift in the definition table types. Our plan was to manually alter the table types to be memory optimized, but only on the largest whales. Introducing schema drift is not ideal, but we made this choice deliberately.

This whole challenge could have been avoided if memory optimized table types didn’t require an In Memory OLTP file group. I get the sense that the memory optimized table types don’t actually use that folder because I noticed something interesting. SQL Server 2019 introduces memory optimized tempdb metadata tables without a memory optimized filegroup! How did they pull that off? I’m a bit jealous. I asked Pam Lahoud and it turns out that the In Memory filegroups are still required for memory optimized table types and will continue to be. It turns out that Microsoft can make certain assumptions about the tempdb metadata tables that they can’t with regular table types. 😟

Some Implementation Surprises

As we implemented our plan, we also encountered some interesting things during testing that might be useful for you if you’re considering In Memory features.

  • The default directory for storing database files should point to a folder that exists on the database server and on any secondary nodes in the same availability group. So if the default location is E:\SQLData then make sure there’s an E drive on every node. SQL Server will need to create an xtp folder in there.
  • When adding the In Memory OLTP file group, the folder that contains it should also exist on all secondary nodes.
  • In SQL Server 2014, I noticed that the addition of the memory optimized file group required up to 4 Gb of space. In SQL Server 2016, I see that that still happens, but the space isn’t taken until the first memory optimized table type I create. That’s also when the xtp folder gets created.
  • Adjusting the table types to be memory optimized was a challenge because we wanted the process to be online. I wrote about how we pulled that off earlier this week in How to Alter User Defined Table Types (Mostly) Online

Success!

Things worked out really really well for us. Our main goal was to avoid tempdb contention and we succeeded there. But there’s an additional performance boost. When you insert into a regular table variable, that data gets written to tempdb’s transaction log. But that’s not the case for memory optimized tables. So even though I really just care about avoiding contention, the boost in performance is significant and measurable and really nice.

In testing we were finally able to push a 96 CPU machine up to 100% CPU on every core and only then did throughput plateau. No tempdb contention in sight.

In production we also saw the same behavior and we were able to sustain over 200K batch requests per second. No tempdb contention in sight.

Those numbers are nice, tempdb contention has been such a thorn in my side for so long, it’s such a relief to squash that issue once and for all. I now get to focus on the next bottleneck and can leave tempdb contention in the past.

Older Posts »

Powered by WordPress