Michael J. Swart

In the first draft of this series, this post didn’t exist. I wanted to show a really simple example of a column switch and include it in the Blue-Green (Details) post. I planned for something simple. But I ran into some hiccups that I though were pretty instructive, so I turned it into the post you see here.

The Plan

For this demo, I wanted to use the WideWorldImporters database. In table Warehouse.ColdRoomTemperatures I wanted to change the column

ColdRoomSensorNumber INT NOT NULL,

into

ColdRoomSensorLabel NVARCHAR(100) NOT NULL,

because maybe we want to track sensors via some serial number or other code.

The Blue-Green plan would be simple:

The Trouble

But nothing is ever easy. Even SQL Server Data Tools (SSDT) gives up when I ask it to do this change with this error dialog:

There’s two things going on here (and one hidden thing):

1. The first two messages point out that a procedure is referencing the column ColdRoomSensorNumber with schemabinding. The reason it’s using schemabinding is because it’s a natively compiled stored procedure. And that tells me that the table Warehouse.ColdRoomTemperatures is an In-Memory table. That’s not all. I noticed another wrinkle. The procedure takes a table-valued parameter whose table type contains a column called ColdRoomSensorLabel. We’re going to have to replace that too. Ugh. Part of me wanted to look for another example.
2. The last message tells me that the table is a system versioned table. So there’s a corresponding archive table where history is maintained. That has to be dealt with too. Luckily Microsoft has a great article on Changing the Schema of a System-Versioned Temporal Table.
3. One last thing to worry about is a index on ColdRoomSensorNumber. That should be replaced with an index on ColdRoomSensorLabel. SSDT didn’t warn me about that because apparently, it can deal with that pretty nicely.

So now my plan becomes:

Blue The original schema

Aqua After the pre-migration scripts are run

An extra step is required here to update the new column and keep the new and old columns in sync.

Green After the switch, we clean up the old objects and our schema change is finished:

Without further ado, here are the scripts:

Pre-Migration (Add Green Objects)

In the following scripts, I’ve omitted the IF EXISTS checks for clarity.

-- Add the four green objects
 
ALTER TABLE Warehouse.ColdRoomTemperatures
ADD ColdRoomSensorLabel NVARCHAR(100) NOT NULL 
    CONSTRAINT DF_Warehouse_ColdRoomTemperatures_ColdRoomSensorLabel DEFAULT '';
GO
 
ALTER TABLE Warehouse.ColdRoomTemperatures
ADD INDEX IX_Warehouse_ColdRoomTemperatures_ColdRoomSensorLabel (ColdRoomSensorLabel);
GO
 
CREATE TYPE Website.SensorDataList_v2 AS TABLE(
    SensorDataListID int IDENTITY(1,1) NOT NULL,
    ColdRoomSensorLabel VARCHAR(100) NULL,
    RecordedWhen datetime2(7) NULL,
    Temperature decimal(18, 2) NULL,
    PRIMARY KEY NONCLUSTERED (SensorDataListID)
)
GO
 
CREATE PROCEDURE Website.RecordColdRoomTemperatures_v2
    @SensorReadings Website.SensorDataList_v2 READONLY
AS
    --straight-forward definition left as exercise for reader
GO

Pre-Migration (Populate and Keep in Sync)

Normally, I would use triggers to keep the new and old column values in sync like this, but you can’t do that with In-Memory tables. So I altered the procedure Website.RecordColdRoomTemperatures to achieve something similar. The only alteration I made is to set the ColdRoomSensorLabel value in the INSERT statement:

ALTER PROCEDURE Website.RecordColdRoomTemperatures
@SensorReadings Website.SensorDataList READONLY
WITH NATIVE_COMPILATION, SCHEMABINDING, EXECUTE AS OWNER
AS
BEGIN ATOMIC WITH
(
    TRANSACTION ISOLATION LEVEL = SNAPSHOT,
    LANGUAGE = N'English'
)
    BEGIN TRY
 
        DECLARE @NumberOfReadings int = (SELECT MAX(SensorDataListID) FROM @SensorReadings);
        DECLARE @Counter int = (SELECT MIN(SensorDataListID) FROM @SensorReadings);
 
        DECLARE @ColdRoomSensorNumber int;
        DECLARE @RecordedWhen datetime2(7);
        DECLARE @Temperature decimal(18,2);
 
        -- note that we cannot use a merge here because multiple readings might exist for each sensor
 
        WHILE @Counter <= @NumberOfReadings
        BEGIN
            SELECT @ColdRoomSensorNumber = ColdRoomSensorNumber,
                   @RecordedWhen = RecordedWhen,
                   @Temperature = Temperature
            FROM @SensorReadings
            WHERE SensorDataListID = @Counter;
 
            UPDATE Warehouse.ColdRoomTemperatures
                SET RecordedWhen = @RecordedWhen,
                    Temperature = @Temperature
            WHERE ColdRoomSensorNumber = @ColdRoomSensorNumber;
 
            IF @@ROWCOUNT = 0
            BEGIN
                INSERT Warehouse.ColdRoomTemperatures
                    (ColdRoomSensorNumber, ColdRoomSensorLabel, RecordedWhen, Temperature)
                VALUES (@ColdRoomSensorNumber, 
                        'HQ-' + CAST(@ColdRoomSensorNumber AS NVARCHAR(50)), 
                        @RecordedWhen, 
                        @Temperature);
            END;
            SET @Counter += 1;
        END;
 
    END TRY
    BEGIN CATCH
        THROW 51000, N'Unable to apply the sensor data', 2;
 
        RETURN 1;
    END CATCH;
END;

That keeps the values in sync for new rows. But now it’s time to update the values for existing rows. In my example, I imagine that the initial label for the sensors are initially: “HQ-1”, “HQ-2”, etc…

UPDATE Warehouse.ColdRoomTemperatures
SET ColdRoomSensorLabel = 'HQ-' + CAST(ColdRoomSensorNumber as nvarchar(50));

Eagle-eyed readers will notice that I haven’t dealt with the history table here. If the history table is large use batching to update it. Or better yet, turn off system versioning and then turn it back on immediately using a new/empty history table (if feasible).

Post-Migration

After a successful switch, the green application is only calling Website.RecordColdRoomTemperatures_v2. It’s time now to clean up. Again, remember that order matters.

DROP PROCEDURE Website.RecordColdRoomTemperatures;
 
DROP TYPE Website.SensorDataList;
 
ALTER TABLE Warehouse.ColdRoomTemperatures
DROP INDEX IX_Warehouse_ColdRoomTemperatures_ColdRoomSensorNumber;
 
ALTER TABLE Warehouse.ColdRoomTemperatures
DROP COLUMN ColdRoomSensorNumber;

Using the Blue-Green deployment method, database changes are decoupled from applications changes. That leaves us with one last challenge to tackle. The schema changes have to be performed while the application is online. It’s true that you can’t always write an online script for every kind of schema change you want.

The challenge of writing online schema changes is essentially a concurrency problem and the guiding principle I follow is: Do whatever you need to do, but avoid excessive blocking.

Locks Are Hot Potatoes

You can’t hold them for long. This applies to schema changes too. Logically if you don’t hold a lock long, you can’t block activity. One exception might be the SCH-M lock which can participate in blocking chains:

SCH-M locks

There are two main kinds of SQL queries. SELECT/INSERT/UPDATE/DELETE statements are examples of Data Manipulation Language (DML). CREATE/ALTER/DROP statements are examples of Data Definition Language (DDL).

With schema changes – DDL – we have the added complexity of the SCH-M lock. It’s a kind of lock you don’t see with DML statements. DML statements take and hold schema stability locks (SCH-S) on the tables they need. This can cause interesting blocking chains between the two types where new queries can’t start until the schema change succeeds:

Some suggestions:

• Don’t rebuild indexes while changing schema
• Rely on the OLTP workload which has many short queries. In an OLTP workload, the lead blocker shouldn’t be a lead blocker for long. Contrast that with an OLAP workload with long-running and overlapping queries. OLAP workloads can’t tolerate changing tables without delays or interruptions.
• When using Enterprise Edition, use ONLINE=ON for indexes. It takes and holds a SCH-M lock only briefly.

Changes to Big Tables

Scripts that change schema are one-time scripts. If the size of the table is less than 50,000 rows, I write a simple script and then move on.

If the table is larger, look for metadata-only changes. For example, these changes are metadata-only changes:

• New columns that are nullable
• New columns with a default value
• Some column alterations with Enterprise Edition (see Altering Text Columns: Only a Metadata Change)

If a table change is not a meta-data change, then it’s a size-of-data change. Then it’s time to get creative. Look for my other post in this series for an example of batching and an example of a column switcheroo.

Pragmatism Example

If you think “good enough” is neither, you may want to skip this section. There are some schema changes that are still very difficult or impossible to write online. With some creativity, we’ve always been able to mitigate these issues with shortcuts and I want to give an example which I think is pretty illustrative.

When a colleague asked for a rowversion column on a humongous table. We avoided that requirement by instead creating a datetime column called LastModifiedDate. Since 2012, new columns with constant default values are online. So we added the column with a constant default, and then changed the default value to something more dynamic:

alter table dbo.MYTABLE
add LastModifiedDate DATETIME NOT NULL 
    CONSTRAINT DF_TABLE_LastModifiedDate DEFAULT '20000101'
 
alter table dbo.MYTABLE
drop CONSTRAINT DF_TABLE_LastModifiedDate;
 
alter table dbo.MYTABLE
add CONSTRAINT DF_TABLE_LastModifiedDate 
      DEFAULT GETUTCDATE() for LastModifiedDate;

It’s a cool situation because it seems like the column has two defaults, one constant default for rows with missing values. And another definition to be used for new rows:

select pc.default_value, d.definition as [default definition]
from sys.system_internals_partitions p
join sys.system_internals_partition_columns pc 
	on p.partition_id = pc.partition_id
join sys.default_constraints d
	on d.parent_object_id = p.object_id
	and d.parent_column_id = pc.partition_column_id
where p.object_id = object_id('MYTABLE')
and pc.partition_column_id = 2
 
/* Gives 
 
default_value  default definition
-------------  ------------------
2000-01-01     (getutcdate())
 
*/

So be creative and pragmatic. Successful 100% online schema changes involve creativity and close collaboration between everyone involved.

Recently I tackled an issue where a DateTime field was getting updated too often. The query looked something like this:

UPDATE dbo.SOMETABLE
SET MyDateTime = GETUTCDATE()
WHERE SomeTableId = @SomeTableId;

So I decided to give up accuracy for concurrency. Specifically, I decided to only update MyDateTime if the existing value was more than a second ago.

First Attempt: Use DATEDIFF With SECOND

My first attempt looked like this:

UPDATE dbo.SOMETABLE
SET MyDateTime = GETUTCDATE()
WHERE SomeTableId = @SomeTableId
AND DATEDIFF(SECOND, GETUTCDATE(), MyDateTime) > 1;

But I came across some problems. I assumed that the DATEDIFF function I wrote worked this way: Subtract the two dates to get a timespan value and then return the number of seconds (rounded somehow) in that timespan.

But that’s not how it works. The docs for DATEDIFF say:

There’s no rounding involved. It just counts the ticks on the clock that are heard during a given timespan.

Check out this timeline. It shows three timespans and the DATEDIFF values that get reported:

But that’s not the behavior I want.

How About DATEDIFF with MILLISECOND?

Using milliseconds gets a little more accurate:

UPDATE dbo.SOMETABLE
SET MyDateTime = GETUTCDATE()
WHERE SomeTableId = @SomeTableId
AND DATEDIFF(MILLISECOND, GETUTCDATE(), MyDateTime) > 1000;

And it would be good for what I need except that DATEDIFF using MILLISECOND will overflow for any timespan over a month. For example,

SELECT DATEDIFF (millisecond, '2017-11-01', '2017-12-01')

gives this error:

Msg 535, Level 16, State 0, Line 1
The datediff function resulted in an overflow. The number of dateparts separating two date/time instances is too large. Try to use datediff with a less precise datepart.

SQL Server 2016 introduced DATEDIFF_BIG to get around this specific problem. But I’m not there yet.

Use DATEADD

I eventually realized that I don’t actually need to measure a timespan. I really just need to answer the question “Does a particular DateTime occur before one second ago?” And I can do that with DATEADD

UPDATE dbo.SOMETABLE
SET MyDateTime = GETUTCDATE()
WHERE SomeTableId = @SomeTableId
AND MyDateTime < DATEADD(SECOND, -1, GETUTCDATE()) ;

Update: Adam Machanic points out another benefit to this syntax. The predicate AND MyDateTime < DATEADD(SECOND, -1, GETUTCDATE()) syntax is SARGable (unlike the DATEDIFF examples). Even though there might not be a supporting index or SQL Server might not choose to use such an index it in this specific case, I prefer this syntax even more.

So How About You?

Do you use DATEDIFF at all? Why? I'd like to hear about what you use it for. Especially if you rely on the datepart boundary crossing behavior.

Devops, Agile and Continuous Delivery

Sometimes I get asked about work and our software development practices. Often these questions use words like agile, devops, or continuous delivery as in “Are you agile?” or “Do you do continuous delivery?”. But these questions rarely have yes or no answers. I always want to answer “It’s a work in progress”.

One of the things I like best about the PASS Summit is the opportunity to talk to people to find out how they do development. After speaking with others, it seems like everyone’s processes are always going to be works-in-progress. But I’ve come to realize that at D2L, we’re pretty far along. Here are just some of the things we do:

Things We Do

• Deployments to production are scheduled and almost completely automatic. This lets us deploy way more often than we used to.
• All code – including procedures and table definitions – are checked in.
• Automatic tests are run on every pull request and merge.
• Safety is a huge concern. We use feature flags and other techniques, but it remains difficult to maintain large complicated systems safely.
• We use blue-green deployments for zero downtime.
• The database layer is an exception to this blue-green technique. So it’s very important to be able to rollback any changes or configuration.

Sardines and Whales

This means we must also support thousands of copies of our main database. They’re used for client sites, test sites, qa sites, or whatever. So that leads to a variety of server configurations that I refer to as sardines and whales:

Look at those sardines. They’re quite happy where they are. The server can handle up to a thousand databases when there’s almost no activity.

But that whale is on a huge server and is extremely busy. Because of the high volume of transactions, we sometimes encounter tempdb contention due to our frequent use of table valued parameters. One technique I’ve been looking forward to evaluating is using memory optimized table types.

Maybe We Can Use In Memory OLTP?

I’m actually not very interested in memory optimized tables. I’m much more interested in the memory optimized table types. Those types can be used for table valued parameters. I can’t tell you how excited I was that it might solve my tempdb pet peeve.

But our dreams for the feature died

We’re leaving the feature behind for a few reasons. There’s an assumption we relied on for the sardine servers: Databases that contain no data and serve no activity should not require significant resources like disk space or memory. However, when we turned on In Memory OLTP by adding the filegroup for the memory-optimized data, we found that the database began consuming memory and disk (about 2 gigabytes of disk per database). This required extra resources for the sardine servers. So for example, 1000 databases * 2Gb = 2Tb for a server that should be empty.

Another reason is that checkpoints began to take longer. Checkpoints are not guaranteed to be quick, but on small systems they take a while which impacts some of our Continuous Integration workflows.

At the PASS Summit, I talked to a Hekaton expert panel. I also talked to a couple people in the Microsoft SQL Server clinic about some of my issues. They all recommended that we upgrade to SQL Server 2016 (which we can’t yet). Maybe I didn’t phrase my questions well, but I didn’t come away with any useful strategy to pursue.

I later talked to a Speaker Idol contestant Brian Carrig (@briancarrig) after hearing him talk briefly about his experiences with In Memory OLTP. He mentioned his own hard-fought lessons with In Memory OLTP including some uncomfortable outages.

The final nail in the coffin, as it were, is that once you turn on In Memory OLTP, you can’t turn it off. Once the In Memory OLTP filegroup is added, it’s there for good. Like I said, safety is a huge concern for us so we’re giving up on the feature for now.

Resurrecting the Feature?

The feature was designed for whales, not sardines. Maybe someday we will try to get those sardine servers to not fail with In Memory OLTP. Until then, the feature goes back on the shelf.

We encountered a CPU issue which took over a month to understand and fix. Since then, it’s been another couple months and so I think it may be time for a debrief.

The cause was identified as a growing number of ghost records that SQL Server would not clean up no matter what. Our fix was ultimately to restart SQL Server.

Symptoms and Data Collection

Here’s what we found.

• At time marked ‘A’ on the graph, we noticed that CPU increased dramatically. It was hard not to notice.
• We used sp_whoisactive to identify a frequent query that was taking a large amount of CPU. That query had never been a problem before. It was a select against a queue table – a table whose purpose was to store data for an implementation of a work queue. This table had a lot of churn: many INSERTS and DELETES. But it was small, never more than 100 rows.
• So next, we ran the query manually in Management Studio. SET STATISTICS IO, TIME ON gave us a surprise. A simple COUNT(*) of the table told us there were 30 rows in the table, but reading it took 800K logical page reads!
• What pages could it be reading? It’s impossible for a table to be that fragmented, it would mean less than one row per page. To look at the physical stats we ran:

  select sum(record_count) as records, sum(ghost_record_count) as ghost_records, sum(version_ghost_record_count) as version_ghost_records from sys.dm_db_index_physical_stats(db_id(), object_id('<table_name>'), default, default, 'detailed') where index_id = 1 and index_level = 0

  select sum(record_count) as records, sum(ghost_record_count) as ghost_records, sum(version_ghost_record_count) as version_ghost_records from sys.dm_db_index_physical_stats(db_id(), object_id('<table_name>'), default, default, 'detailed') where index_id = 1 and index_level = 0

  And that gave us these results:

  
  Interesting. The ghost records that remain are version_ghost_records, not ghost_records. Which sounds like we’re using some sort of snapshot isolation (which we’re not), or online reindexing (which we are), or something else that uses row versions.

• Over time, those version_ghost_records would constantly accumulate. This ghost record accumulation was also present in all other tables, but it didn’t hurt as much as the queue table which had the most frequent deletes.

Mitigation – Rebuild Indexes

Does an index rebuild clean these up? In this case, yes. An index rebuild reduced the number of version ghost records for the table. At the time marked ‘B’ in the timeline, we saw that an index rebuild cleaned up these records and restored performance. But only temporarily. The version_ghost_records continued to accumulate gradually.

At time ‘C’ in the timeline, we created a job that ran every 15 minutes to rebuild the index. This restored performance to acceptable levels.

More investigation online

Kendra Little – Why Is My Transaction Log Growing In My Availability Group?
This is a great video. Kendra describes a similar problem. Long running queries on secondary replicas can impact the primary server in an Availability Group (AG). But we found no long running queries on any replica. We looked using sp_whoisactive and DBCC OPENTRAN. We didn’t see any open transactions anywhere.

Amit Banerjee – Chasing the Ghost Cleanup in an Availability Group
Amit also mentions that log truncation would be prevented in the case of a long-running query on a replica. However, in our case, log truncation was occurring.

Uwe Ricken – Read Committed Snapshot Isolation and high number of version_ghost_record_count
Uwe Ricken also blogged recently about a growing number of version_ghost_records. He talked about looking for open transactions that use one of the snapshot isolation levels. Unfortunately it didn’t apply to our case.

Bitemo Erik Gergely – A very slow SQL Server
Another example of a long running query keeping version_ghost_records around.

dba.stackexchange – GHOST_CLEANUP Lots of IO and CPU Usage
This stackexchange question also describes a problem with lots of CPU and an inefficient, ineffective ghost cleanup task for databases in an AG. There’s an accepted answer there, but it’s not really a solution.

Calling Microsoft Support

So we called Microsoft support. We didn’t really get anywhere. We spoke with many people over a month. We generated memory dumps, PSSDiag sessions and we conducted a couple screen sharing sessions. Everyone was equally stumped.

After much time and many diagnostic experiments. Here’s what we did find.

• Availability Groups with readable secondaries are necessary (but not sufficient) to see the problem. This is where the version_ghost_records come from in the first place. Readable secondaries make use of the version store.
• We ran an extended event histogram on the ghost_cleanup event. There was a ridiculous amount of events. Like millions per minute, but they weren’t actually cleaning up anything:

  CREATE EVENT SESSION ghostbusters ON SERVER ADD EVENT sqlserver.ghost_cleanup( ACTION( sqlserver.database_id ) ) ADD TARGET package0.histogram( SET filtering_event_name=N'sqlserver.ghost_cleanup', source=N'sqlserver.database_id' )

  CREATE EVENT SESSION ghostbusters ON SERVER ADD EVENT sqlserver.ghost_cleanup( ACTION( sqlserver.database_id ) ) ADD TARGET package0.histogram( SET filtering_event_name=N'sqlserver.ghost_cleanup', source=N'sqlserver.database_id' )

• Microsoft let me down. They couldn’t figure out what was wrong. Each new person on the case had to be convinced that there were no open transactions. They couldn’t reproduce our problem. And on our server, we couldn’t avoid the problem.

Resolution

Ultimately the time came for an unrelated maintenance task. We had to do a rolling upgrade for some hardware driver update. We manually failed over the Availability Group and after that, no more problem!

It’s satisfying and not satisfying. Satisfying because the problem went away for us and we haven’t seen it since. After the amount of time I spent on this, I’m happy to leave this problem in the past.

But it’s not satisfying because we didn’t crack the mystery. And restarting SQL Server is an extreme solution for a problem associated with an “Always On” feature.

If you’re in the same boat as I was, go through the links in this post. Understand your environment. Look for long running queries on all replicas. And when you’ve exhausted other solutions, mitigate with frequent index rebuilds and go ahead and restart SQL Server when you can.