Update and Correction:

This blog was originally posted on February 20. Since then I read other articles that suggested different behavior with the Halloween Problem. I contacted Paul White, who informed me that the WideWorldImporters database uses compatibility level 130 (SQL Server 2016) by default. So, I tested on a SQL Server 2019 instance but was probably seeing an issue addressed in later updates.

I tested again at compatibility level 150 and saw a different execution plan which led to different conclusions.

I’ve left the majority of the post unchanged, but I’m adding an addendum section, and updating the summary and its conclusions. So, make sure you read those sections for the corrections.

Original Post:

I find myself talking about the Halloween Problem a lot and wanted to fill in some more details on the subject. In short, the Halloween Problem is a case where an INSERT\UPDATE\DELETE\MERGE operates on a row more than once, or tries to and fails. In the first recorded case, an UPDATE changed multiple rows in the table more than once.

So let’s take a look at an example using a publicly available database, WideWorldImporters.

A Halloween Problem example

Here’s a simple update procedure. We’re going to update the quantity for an item in the Sales.OrderLines table:

CREATE OR ALTER PROCEDURE Sales.OrderLines_UpdateQuantity
	@OrderID INT,
	@StockItemID INT,
	@Quantity INT
	WITH EXECUTE AS OWNER
AS
BEGIN
	SET NOCOUNT ON;
	SET XACT_ABORT ON;

	UPDATE sol
	SET
		sol.Quantity = @Quantity,
		sol.PickedQuantity = @Quantity
	FROM Sales.OrderLines sol
	WHERE
		sol.OrderID = @OrderID
		AND sol.StockItemID = @StockItemID;
		-- AND sol.Quantity <> @Quantity;
END;
GO

You may notice the commented line. In one description of the Halloween Problem I heard\read, it was suggested that if we try to SET something that is in our WHERE clause the problem is likely to occur. Or rather, SQL Server will see the possibility of the problem and add protections to our execution plan to prevent it.

First, let’s test without that line, and see what our execution plan tells us.

EXEC Sales.OrderLines_UpdateQuantity
	@OrderID = 5,
	@StockItemID = 155,
	@Quantity = 21;
GO
Note the eager spool

The eager spool between our index reads and clustered index update shows that SQL Server added Halloween protections to prevent the problem. The problem is prevented by separating the read phase of the query from the write phase.

This usually involves a blocking operator. Most often this is an eager spool, but if there is another blocking operator in the plan like a sort or hash match, that blocking operator may remove the need for a separate spool.

The Halloween Problem would occur if a query is running in row mode and as rows are still being read, rows are being updated and moved in an index. This allows the read operation to potentially read the updated row again and operate on it again. The index movement is key in this scenario.

But with a blocking operator between the read operation and the write, we force all the reads to complete first. This gives us a complete, distinct list of rows to be updated (in this example) before we get to the clustered index update, so it isn’t possible to update the same row twice.

So, how does index movement come into play here? We are updating the Quantity and PickedQuantity columns in our UPDATE statement. Both fields are key columns in the only columnstore index on the table, NCCX_Sales_OrderLines.

CREATE NONCLUSTERED COLUMNSTORE INDEX [NCCX_Sales_OrderLines] ON [Sales].[OrderLines]
(
[OrderID],
[StockItemID],
[Description],
[Quantity],
[UnitPrice],
[PickedQuantity]
)WITH (DROP_EXISTING = OFF, COMPRESSION_DELAY = 0) ON [USERDATA];
GO

So when we update these columns, the affected rows will move in that index. If the row moves, that means a read operation could continue reading and find the same row again, returning it as a part of its result set a second time.

Interestingly, we aren’t reading from the columnstore index in the plan provided. Since that’s the only index containing these columns as key values, it’s the only index where the rows should move. In this case, our read operators shouldn’t encounter updated rows a second time, since they use the FK_Sales_OrderLines_OrderID (with a key lookup against PK_Sales_OrderLines).

I wonder if SQL Server decided the Halloween protections were needed before it decided which index it would use for the read.

Removing the index

Either way, if we dropped the NCCX_Sales_OrderLines index, we should see a plan without an eager spool between the read operators and the update operator.

IF EXISTS(
	SELECT 1
	FROM sys.indexes si
	WHERE
		si.name = 'NCCX_Sales_OrderLines'
)
BEGIN
	DROP INDEX [NCCX_Sales_OrderLines]
		ON Sales.OrderLines;
END;
GO

With the index removed, let’s look at the new plan.

Unspooled

We’ve lost the extra steps to the left of the clustered index update operator to update the columnstore index, and we have also lost the eager spool between the read operators and the update operator. This shows without the index movement, Halloween protections are no longer needed.

Performance impact of protections

Let’s look at the data from Query Store to see how big the difference is between the two execution plans.

I ran a simple query against the same OrderID in Sales.OrderLines before running the procedure before and after the index change to get the data into the cache (because cold cache issues were making a large difference). I also ran the procedure 10 times to try to average out our results in case any odd wait types were seen.

80 microseconds versus 46 microseconds. Blazing fast in both cases with the data already cached, but the plan with Halloween protections took 74% longer. Unsure if the update to a columnstore index is significantly more expensive than that of a rowstore index. Perhaps we should test this again without columnstore complicating the issue.

Speaking in general, I would expect a bigger difference in a query affecting more rows. For a query that only returns 3 rows from the first index seek, the delay caused by the spool would be very small. But imagine if we have a query that reads tens or hundreds of thousands of rows before performing its write operation.

Normally such a query would be passing rows it has read up to the join and update operators while it is continuing to read. Those operations would be happening on different threads in parallel.1

If we are being protected from the Halloween Problem, the eager spool will not return any rows to the operations above it (like the clustered index update) until all rows have been read. So the writes cannot start until much later, and the more rows being read the more considerable the delay.

Nonclustered indexes?

If you noticed the “+3 non-clustered indexes” banner in one of the plans above, that’s indicating the nonclustered indexes updated when we updated the clustered index. This is more obvious in Plan Explorer than in the plans as shown in SQL Server Management Studio. So, I wanted to point that out in case the visual was confusing to anyone.

But this raises another question. If we are updating those indexes, why don’t they cause the Halloween protections to be used?

That is because the quantity columns are present in those indexes only as included columns. Changes to those columns won’t affect where the row sorts, but the values still need to be updated.

Rowstore testing

So, let’s see how this looks with a rowstore index. Here’s a second procedure, similar to the first but also updating PickingCompletedWhen.

CREATE OR ALTER PROCEDURE Sales.OrderLines_UpdateQuantityWhen
	@OrderID INT,
	@StockItemID INT,
	@Quantity INT
	WITH EXECUTE AS OWNER
AS
BEGIN
	SET NOCOUNT ON;
	SET XACT_ABORT ON;

	UPDATE sol
	SET
		sol.Quantity = @Quantity,
		sol.PickedQuantity = @Quantity,
		sol.PickingCompletedWhen = GETUTCDATE()
	FROM Sales.OrderLines sol
	WHERE
		sol.OrderID = @OrderID
		AND sol.StockItemID = @StockItemID
		AND sol.PickingCompletedWhen < GETUTCDATE();
END;
GO

Initially, no index uses PickingCompletedWhen. So if we execute the procedure as is, we shouldn’t see the tell-tale eager spool.

EXEC Sales.OrderLines_UpdateQuantityWhen
	@OrderID = 5,
	@StockItemID = 155,
	@Quantity = 21;
GO

This plans is what we’d expect. If we add an index, how does this change the plan and how does this change the performance?

IF NOT EXISTS(
	SELECT 1
	FROM sys.indexes si
	WHERE
		si.name = 'IX_OrderLines_OrderID_StockItemID_PickingCompletedWhen'
)
BEGIN
	CREATE INDEX IX_OrderLines_OrderID_StockItemID_PickingCompletedWhen
		ON Sales.OrderLines (OrderID, StockItemID, PickingCompletedWhen);
END;
GO

Here, we see the eager spool implementing the Halloween protections again, but between the index seek and the key lookup. Note that the new index is the one we are using for the index seek. The clustered index update now indicates it is updating 4 nonclustered indexes, including the new index.

So, is the performance difference as stark as it was with the columnstore index?

So, 52 µs vs 118 µs. The query took about ~126% longer when the Halloween protections were present. More than we saw with the columnstore index, which is surprising. Perhaps it is relevant that we are updating a third field. It almost feels like the observer effect at this scale.

Addendum

So, to correct things here, let’s go back to the first procedure.

CREATE OR ALTER PROCEDURE Sales.OrderLines_UpdateQuantity
	@OrderID INT,
	@StockItemID INT,
	@Quantity INT
	WITH EXECUTE AS OWNER
AS
BEGIN
	SET NOCOUNT ON;
	SET XACT_ABORT ON;

	UPDATE sol
	SET
		sol.Quantity = @Quantity,
		sol.PickedQuantity = @Quantity
	FROM Sales.OrderLines sol
	WHERE
		sol.OrderID = @OrderID
		AND sol.StockItemID = @StockItemID;
		-- AND sol.Quantity <> @Quantity;
END;
GO

If I run this procedure again with the restored database and no other changes besides updating the compatibility level to 150, I see the following execution plan:

So, we have no eager spool, which means the Halloween Problem isn’t a problem now.

Previously, there was a spool between the index seek and lookup and the clustered index update. The only index using any of the updated fields as a key value was the columnstore index. This suggested that the optimizer will use Halloween protections if any index uses the updated fields as a key value because the rows would be moved in that index.

This new plan disproves that because the optimizer no longer uses the protections with the later compatibility level. And the columnstore index (NCCX_Sales_OrderLines) is still present (as you can see if you hover over the clustered index update operator).

As for the second procedure, I see the Halloween protections even without the index I added in my example. Without that index, the query originally used the FK_Sales_OrderLines_OrderID index to seek the rows in question. At the higher compatibility level, the IX_Sales_OrderLines_Perf_20160301_02 index is used, which is keyed on (StockItemID, PickingCompletedWhen).

So, the Halloween protections are used because we read from an index keyed on one of the updated fields, and rows being updated will potentially move in that index.

We’ve seen the Halloween protections when using nonclustered indexes so far, but what if we are using the clustered index for the read?

I wrote a quick procedure to change the OrderLineID, which is the only column in the clustered primary key for this table. And this matches expectations; we see the eager spool between the clustered index seek and the update operator.

Summary

Hopefully, the addendum corrects the matter while keeping things clear. I’m updating one of the bullet points below, as well.

It seems there are only two criteria for the protections against the Halloween Problem to be used for an UPDATE query:

  1. The object being updated must also be in the query.
  2. A column being updated must be a key column in at least one index on the table. One of the updated columns must be a key value in the index used for the read portion of the query, so that the rows may move in that index.

For other statements, the setup is more complex. I find the UPDATE statement is the most straightforward example of the Halloween Problem. But you can see the protections in place if you query from a table as part of an INSERT or DELETE (or MERGE) where you change that same table.

And if we see Halloween protections in the plan for a query, we could change the offending index or the query to change the behavior.

Or we could use the manual Halloween technique, which I will discuss next time.

Thanks again to Paul White for pointing out the compatibility level; I doubt that would ever have occurred to me.

Please contact me if you have any questions or comments. I’ve updated my social media links above to include Counter.Social and Mastodon. We’ll see if there is more #sqlfamily activity on those platforms going forward.

Footnotes

1: Not the type of parallelism we typically think of with SQL Server. Parallelism is typically when a given operation, like an index scan, is expected to process many rows, and SQL Server dedicates multiple threads to that operator or group of operators. In this case, I say parallel because different operators (the index seek, nested loops join, and clustered index update) are all processing rows at the same time, one row at a time.

Tempdb contention has long been an issue in SQL Server, and there are many blogs on the issue already. But I wanted to add one more mainly to highlight the improvements in recent versions of SQL Server

Tempdb contention is most often discussed in as relating to the creation of temp tables (and other objects) in tempdb. If you are experiencing this you will see PAGELATCH_EX or PAGELATCH_SH waits, frequently with wait resources like 2:1:1 or 2:1:3. This indicates contention in database 2 (tempdb), page 1 (the first data file in tempdb), and one of the PFS, GAM, or SGAM pages (which are pages 1, 2, and 3 respectively). Tempdb files of sufficient size will have additional PFS, GAM, and SGAM pages at higher page numbers, but 1 and 3 are the pages most often referenced.

Temp tables aren’t the only objects being created in tempdb. Table variables are as well unless they are memory-optimized. There are also worktables for sorts, spools, and cursors. Hash operations can spill to disk and are written into tempdb. Row versions are written into tempdb for things like read committed snapshot isolation, and triggers make use of row versioning as well. For more details, check out this excellent post by David Pless.

Before recent releases, there were three main suggestions for reducing tempdb contention.

  • Trace flags (1118 and 1117)
  • More tempdb files
  • Create fewer objects in tempdb

Honestly, I don’t think the third was even included in a lot of the blogs on the subject, and it is very important. Many of the actions that use tempdb can’t be avoided, but I tend to use memory-optimized table variables instead of temp tables the vast majority of the time.

In one case a few years ago, I replaced the memory-optimized table variables in one very frequently executed stored procedure with temp tables to see if using temp tables would result in better execution plans. This procedure was executed about 300 million times per day across several SQL Server instances using similar databases, and the procedure used 4 temp tables. The plans didn’t matter; creating 1.2 billion more temp tables per day added far too much tempdb contention.

But the main point of this post is to help everyone catch up on the topic, and see how more recent versions of SQL Server improve on this issue.

Improvements in SQL Server 2016

SQL Server 2016 introduced several improvements that help reduce tempdb contention.

The most obvious is that setup will create multiple files by default, one per logical server up to eight. That bakes in one of the main recommendations for reducing tempdb contention, so it’s a welcome improvement.

There are also behavior changes that include the behavior of trace flags 1117 and 1118. All tempdb data files grow at the same time and by the same amount by default, which removes the need for trace flag 1117. And all tempdb allocations use uniform extents instead of mixed extents, removing the need for trace flag 1118.

So, that’s another recommendation for reducing tempdb contention already in place.

Several other changes also improve caching (reducing page latch and metadata contention), reduce the logging for tempdb operation, and reduce the usage of update locks.
For the full list, check here.

Improvements in SQL Server 2019

The big change here is the introduction of memory-optimized tempdb metadata. The documentation here says that this change (which is not enabled by default, you will need to run an ALTER SERVER CONFIGURATION statement and restart) “effectively removes” the bottleneck from tempdb metadata contention.

However, this post by Marisa Mathews indicates the memory-optimized tempdb metadata improvement in SQL Server 2019 removed most contention in PFS pages caused by concurrent updates. This is done by allowing the updates to occur with a shared latch (see the “Concurrent PFS updates” entry here).

Tempdb contention seen in sp_WhoIsActive output

One thing I would point out is that the metadata is being optimized here; the temp tables you create are not memory-optimized and will still be written to the storage under tempdb as usual.

Improvements in SQL Server 2022

The post above also indicates that SQL Server 2022 reduces contention in the GAM and SGAM pages by allowing these pages to be updated with a shared lock rather than an update log.

The issue with the PFS, GAM, and SGAM pages has always been the need for an exclusive latch on those pages when an allocation takes place. If 20 threads are trying to create a temp table, 19 of them get to wait. The suggestion to add more data files to tempdb was a way to get around access to these pages being serialized; adding more files gives you more of these pages to spread the allocation operations across.

In Summary

The gist is that tempdb contention has been nearly eliminated in SQL Server 2022. There are still several other actions that use tempdb, and you may see contention if have a niche workload or use a lot of worktables.

Hopefully, this post will help you decide if it’s time for an upgrade. If you have been seeing tempdb contention on these common pages, the latest release should be a major improvement.

Feel free to contact me with any questions or let me know of any suggestions you may have for a post.

Wanted to point out a few more good articles and a video on the subject that you may enjoy.

I’ve discussed the other two join types, so what is the niche for the third?

Before we get into how it works and what my experience is I want to mention a response to my last blog, because it leads into our topic.

Addendum on Hash Match Joins

My last blog post was on hash match joins, and Kevin Feasel had a response on his blog.

Hash matches aren’t inefficient; they are the best way to join large result sets together. The caveat is that you have a large result set, and that itself may not be optimal. Should it be returning this many rows? Have you included all the filters you can? Are you returning columns you don’t need?

Jared Poche

I might throw in one caveat about hash match joins and being the best performers for two really large datasets joining together: merge join can be more efficient so long as both sets are guaranteed to be ordered in the same way without an explicit sort operator. That last clause is usually the kicker.

Kevin Feasel, Curated SQL

And he is quite correct. Nested loops perform better than hash match with smaller result sets, and hash match performs better on large result sets.

Merge joins should be more efficient than both when the two sources are sorted in the same order. So merge joins can be great, but the caveat is that you will rarely have two sources that are already sorted in the same order. So if you were looking for the tldr version of this blog, this paragraph is it.

How Merge Joins Operate

Merge joins traverse both inputs once, advancing a row at a time and comparing the values from each input. Since they are in the same order, this is very efficient. We don’t have to pay the cost to create a hash table, and we don’t have the much larger number of index seeks nested loops would encounter.

The process flows like this:

  1. Compare the current values from each data source.
  2. If they match, add the joined row to the result set, and get the next value from both sources.
  3. If not, get the next row from the data source with the lower sorted value.
  4. If there are no more rows from either source, the operation ends.
  5. Otherwise, return to step 1 with the new input.

At this point, I would create a great visual for this, but one already exists. So let me refer you a post by Bert Wagner. The video has a great visualization of the process

Input Independence

I find nested loops is probably the easiest join to understand, so I want to draw a distinction here. Using nested loops, we would get a row from the first source then seek the index against the second to get all rows related to the row from the first source. So, our ability to seek from the second depends on the first.

A merge join seeks from both independently, taking in rows and comparing them in order. So in addition to the requirement (with exception) that the sources have to be in the same order, we need a filter we can use for each source. The ON clause does not give us the filter for the second table, we need something else.

Here’s an example query and plan:

USE WideWorldImporters
GO
SELECT 
	inv.InvoiceID,
	invl.InvoiceLineID
FROM Sales.Invoices inv
INNER JOIN Sales.InvoiceLines invl
	ON invl.InvoiceID = inv.InvoiceID
WHERE
	inv.InvoiceID < 50;
GO

Both Invoices and InvoiceLines have indexes based on InvoiceID, so the data should already be in order. So this should be a good case for a merge (the nested loops below is because of the key lookup on InvoiceLines). But SQL Server’s optimizer still chose nested loops for this query.

I can hint it to get the behavior I expected, and that plan is below.

The estimates are way off for the Invoices table, which is odd considering we are seeking on the primary key’s only column; one would expect that estimate to be more accurate. But this estimate causes the cost for the seek against Invoices to be more expensive, so the optimizer chose the other plan. It makes sense.

I updated the statistics, and a different plan was chosen. One with a hash match.

???

In that case, the difference in cost was directly the cost of the join operator itself; the cost of the merge join operator was 3x the cost of the hash match operator.

Even if the merge is more efficient, it seems it’s being estimated as being more costly, and specifically for CPU cost. You’re likely to see merge joins much less often than the other two types because of the sort requirement; how it is estimated may also be a factor.

About that sort

The first several times I saw a merge join in an execution plan, the merge was basically the problem with the query. It gave me the impression at the time that merge joins aren’t great in general. But in those cases, the execution plan had a sort after one of the index operations and before the join. Sure, the merge join requires that the two sources be sorted in the same order, but SQL Server could always use a sort operator (expensive as they are) to make that an option.

This seems like an odd choice to make, so let’s consider the following query:

USE WideWorldImporters
GO
SELECT *
FROM Sales.Invoices inv
INNER JOIN Sales.InvoiceLines invl
	ON invl.InvoiceID = inv.InvoiceID
WHERE
	inv.InvoiceDate < DATEADD(month, -12, getutcdate());
GO

So, this query does a merge join between the two, but there is a sort on the second input. We scan the index, then sort the data to match the other import before we perform the actual join. A sort operator is going to be a large cost to add into our execution plan, so why did the optimizer choose this plan?

This is a bad query, and the optimizer is trying to create a good plan for it. This may explain many other situations where I have seen a sorted merge. The query is joining the two tables on InvoiceID, and the only filter is on Invoices.InvoiceDate. There is no index on Invoices.InvoiceDate, so it’s a given we’ll scan that table.

If this query used nested loops, we could use the InvoiceID for each record from Invoices to seek a useful index against InvoiceLines, but that would mean we perform 151,578 seeks against that table.

A merge join, even if we have to sort the results from the table, would allow us to perform one index operation instead. But a merge join has to seek independently from the other source, and no other filter is available. So we perform an index scan against the second table as well.

This is probably the best among poor options. To really improve this query, you’d need to add an index or change the WHERE clause.

It took some time for me to realize why I most often saw merge joins in poor execution plans; I wasn’t seeing all the plans using them that perform well. If you are troubleshooting a high CPU situation, when you find the cause you’ll likely be looking at bad plan. We don’t tend to look for the best performing query on the server, do we?

So, if merge join is more efficient than the other two join types in general, we are less likely to be looking at queries where it is being used effectively.

Summary

Hopefully I’ll be getting back to a more regular schedule for the blog. There’s been a number of distractions (an estate sale, mice, etc), but life has been more calm of late (mercifully).

I spoke at the two PASS Summit virtual events over the last two years, and this year I am happy to be presenting in person at PASS Data Community SUMMIT for the first time. So if you are interested in how you can use memory-optimized table variables to improve performance on your system, look out for that session.