Seeks seeks seeks… The number of the beast. Seeks are cool though. I like seeks. Sometimes, at least. Perhaps some seeks are cooler than others. I just said seeks six times, now we’re ready to go.
Let’s begin by creating a temporary table that holds values from 1 to 10 000.
CREATE TABLE #T(id integer PRIMARY KEY CLUSTERED);
WITH n(n) AS
(SELECT 1 UNION ALL SELECT n+1 FROM n WHERE n < 10000)
INSERT INTO #T(id)
SELECT n FROM n ORDER BY n
OPTION (MAXRECURSION 10000);
Now, let’s assume that we need to retrieve a specific range of values – e.g. between 500 and 563. There are a couple of ways to do that, that’s not exactly rocket science, but let’s consider the following two approaches:
Using the BETWEEN operator:
SELECT
Id
FROM
#X
WHERE
Id BETWEEN 500 AND 563
Using the IN operator:
SELECT
Id
FROM
#X
WHERE
Id in
(
500,501,502,503,504,505,506,507,508,509,510,511,512,
513,514,515,516,517,518,519,520,521,522,523,524,525,
526,527,528,529,530,531,532,533,534,535,536,537,538,
539,540,541,542,543,544,545,546,547,548,549,550,551,
552,553,554,555,556,557,558,559,560,561,562,563
)
Uh, truly fascinating.
But, behind these boring and functional equivalent queries, there is a dirty seekret – Wow, I’m on fire today. Both result in a Clustered Index Seek, but one is sus. Discuss!
That’s quite a difference. The second query results in 64 times more logical reads. But why?
One Seek To Rule Them All
In the execution plan of the query that uses the BETWEEN operator, it’s possible to notice that there’s only one seek operation, which range starts at 500 and then traverses the leaf level of the index, until it finds a row beyond the range limit – 563 in this case.
The Meeseeks Box
The seek #2, in the execution plan of the query that uses the IN operator is rather different:
The list continues down to a total of 64 individual seeks.
The underlying reason for the 128 logical reads is related to the size of our table and consequently, the size of the index. Since it is large enough to need a separate root page, each one of the 64 seeks will imply 2 logical reads.
But I digress, the bottom line here is that Seek #2 is the impostor. It is not actually a seek, but a whole bunch of them.
Sargability is a funny word that’s used to characterize the ability of the engine to take advantage of and index, in order to speed up the execution of a query. The legend says that it is derived from a contraction of Search ARGument ABLE.
Usually, when some writes a query, the goal it’s to retrieve the desired information, as fast as possible. What’s interesting is that “as fast as possible” can either refer to “execution time” or “time I will spend writing this query” and we all know which one is, most of the times. As a rule of thumb, queries must be written in the simplest way possible, being the underlying reason that if it’s easy for you to read, it will also be easier for the Optimizer to read and process it.
But that’s not always true, for example, we use functions because they save us effort and improve query readability, but typically they don’t make Optimizer’s life easier. Au contraire, because they tend to make predicates non-sargable. On other occasions, bighearted developers decide to allow the customer to search on every damn existing column, using LIKE with leading and trailing % wildcards and that’s also a sargable capital sin, as we’re going to look further.
Non-Sargable Predicates
So, let’s talk a bit more about bad ideas. Not those that start with “Hold my beer and watch this”, but instead “It runs fast on my machine, let’s take it to Production”.
Non-sargable predicates have a negative impact on query performance, e.g. by causing poor cardinality estimates or inability to perform index seeks, which consequently will lead to inappropriate plan choices and increased CPU consumption.
The following list include some of the most common patterns that kill predicate sargability. We’ll very briefly overview them and posteriorly focus on one specific case.
ColumnX = Function(ColumnX): As already pointed out, functions can be naughty. One might be tempted to use functions because they encapsulate logic in a very practical way to us humans, but this often implies performance loss. The query presented below, where we extract the year from a date column, comparing it against a specific value, represents one of those cases. SQL Server won’t be able to use any index efficiently, even if we have an index on that field. We’ll shortly discuss why this happens.
SELECT COUNT(1) FROM Matches WHERE YEAR(Date) = 1999
ColumnX = ColumnY + ColumnZ: A similar behaviour is observed when we use the result of arithmetic operations between columns to retrieve results of a particular column. It’s important to stress that a predicate containing an arithmetic operation between a column and a constant literal value would also be non-sargable.
SELECT *
FROM Matches
WHERE FullTimeGoals = HalfTimeHomeGoals + HalfTimeAwayGoals
@Parameter1 IS NULLORColumnX = @Parameter1: The heart of that single query where search dreams come true and someone got really proud of it. It’s possible to search for everything. Slowly. In runtime SQL Server won’t be able to figure this out, hence the non-sargability. There are a couple of ways to get around this, being one of them to add OPTION(RECOMPILE). This hint enables Parameter Embedding Optimization, which gives the Parser the ability to replace query parameters with literal constant values during query parsing and enables further complex simplifications, like removing contradictions. Nevertheless, as this hint forces the recompilation of a statement, it introduces some overhead that needs to be taken into account.
CREATE OR ALTER PROCEDURE SearchNonFriendlyMatch
(
@Hometeam INT = NULL,
@Awayteam INT = NULL,
@Date DATE = NULL,
@FTG TINYINT = NULL
)AS
BEGIN
SELECT
HomeTeam,AwayTeam,Date, FullTimeGoals
FROM
Matches
WHERE
League IS NOT NULL AND
(@HomeTeam IS NULL OR HomeTeam = @HomeTeam) OR
(@AwayTeam IS NULL OR HomeTeam = @HomeTeam) OR
(@Date IS NULL OR [Date] = @Date) OR
(@FTG IS NULL OR FullTimeGoals >= @FTG)
END
ColumnXLIKE ‘%Make This Query Go Slower Please%’
When we have this wildcard prefixing and suffixing a string, we’re telling SQL Server to retrieve all records that contain the specified text, regardless of what comes before or after it.
SELECT TOP 5 * FROM Teams WHERE Name LIKE '%ar%'
The execution of the previous query returns these results:
Id
Country
Name
7
Portugal
SC Beira-Mar
8
England
Cardiff City FC
23
England
Arsenal FC
28
Spain
FC Barcelona
68
Italy
Parma
But why is this bad for performance?
Let’s put ourselves in SQL Server shoes and pretend that we’ve received a list containing names of football teams, ordered alphabetically. We need to count how many of these names contain the letters “ar”. We roll up our sleeves and start analyzing all characters from the first record, then the second, then the third, and so on, until we reach the last item in the list. We had no alternative but to scan through all values, simply because the order of this list wasn’t helpful at all. Recalling the table above it’s easy to understand why: the substring can be everywhere, there is no shortcut, no possibility to discard any subset of these names. SQL Server will have the same exact struggle scanning through all the index represented by our list.
Now, let’s execute a similar query:
SELECT COUNT(1) FROM Teams WHERE t.Name LIKE '%Benf%'
And look into its execution plan:
This is a pretty small table but it’s enough to prove the point. This query ran in about 600 miliseconds and IO statistics show 1 Scan Count and 1226 logical reads. The cost is almost entirely in the Index Scan, that was chosen even thought SQL Server estimated just 9 rows to be retrieved. Well, that wasn’t really a choice, in this case an Index Scan was made because the non-sargable search prevented SQL Server from using an Index Seek, as we already know.
With that being said, let’s remove the leading wildcard and execute the query:
SELECT COUNT(1) FROM Teams t WHERE t.Name LIKE 'Benf%'
And this time, it is possible to confirm by the execution plan that the query was much more efficient using the Index Seek.
It has executed in virtually no time and IO statistics show 1 Scan Count and 3 logical reads.
If we return to our ordered list, it’s very easy to understand why the query executed so much faster: Now we can benefit from that alphabetical order, going directly to the letter B, looking through a few lines and quickly return the row that matches the search.
What Should I Do, Then?
We saw that the sargable version of our query ran in virtually no time, while the non-sargable version took about 600ms and even though proportionally this gap is huge, users don’t complain about half a second queries. Nevertheless, the real problem with this type of queries is they won’t scale. When you try to run this query against a table with millions of rows instead of a couple thousand, it will be much more painful. You’ve been warned.
But what can we do to avoid it?
The first and obvious option would be to analyze if that leading wildcard is really needed. If you can simply get rid of it, problem solved! If not, there’s still hope…
One way to improve these searches is by using Full Text Search (FTS). FTS is able to quickly find all matches for a term search in a table, because it doesn’t have to scan rows. Instead it takes advantage of text indexes that store positional information for all terms found in the columns where we create the text index on. Instead of using LIKE operator, FTS enables the use of CONTAINS/CONTAINSTABLE or FREETEXT/FREETEXTTABLE, as exemplified below:
SELECT COUNT(1) FROM Teams t WHERE CONTAINS((t.Name), @Name)
FTS has a few limitations and requires some setup effort, but I often work with it and it can be pretty awesome.
Additionally, there are newer, hotter technologies designed to solve these issues. For example ElasticSearch, that allows fast text searches in a very scalable way. It also allows pretty complex searches but, to me, the nicest advantage comparing to Full-Text Search is the possibility to take the heavy processing completely outside SQL Server. CPU cores in SQL Server can be very expensive and should be treated with respect, my boss says.
Now you now why you should avoid queries LIKE ‘%these%’ and you have some ideias to improve them. Hope this was helpful to you 🖖
Imagine that you live in a world where you need an applicational user with god-like powers because, in fact, it needs to. If you want to make sure that applicational users are only used by the application you’ll have to create strong passwords, keep’em secret, change’em often, all that good stuff, sure. But you know what they say – when there’s a will, there’s a way – and there are people who simply like to leave in Santa’s naughty list. So today we’ll talk about an old trick that can be used to help prevent logins from a user, under specific cirumstances – Logon Triggers:
“Logon triggers fire stored procedures in response to a LOGON event. This event is raised when a user session is established with an instance of SQL Server. Logon triggers fire after the authentication phase of logging in finishes, but before the user session is actually established. Therefore, all messages originating inside the trigger that would typically reach the user, such as error messages and messages from the PRINT statement, are diverted to the SQL Server error log. Logon triggers do not fire if authentication fails.”
Firstly, we’ll try to deny any SSMS login attempt with the user “appUser”. To accomplish that we’ll create a new trigger, as presented below:
CREATE OR ALTER TRIGGER DenySSMSLogin
ON ALL SERVER
FOR LOGON
AS
BEGIN
DECLARE @AppName varchar(max)
DECLARE @LoginName sysname
DECLARE @LoginType sysname
SET @AppName = APP_NAME()
SET @LoginName = ORIGINAL_LOGIN()
IF(@LoginName = 'appUser' AND @AppName LIKE 'Microsoft SQL Server Management Studio%')
BEGIN
RAISERROR('Call the Federales',16,1)
ROLLBACK;
END
END
Now, we won’t be able to login this user on SSMS, as the following error message indicates:
Most importantly, queries done by this user through the application’s services will be processed as intended. Now just sit and wait for people to start asking what’s happening and then beat them with the nearest heavy object, calmly explain them why this is beeing done.
Ok, so now I have to come clean. The bad news is that I lied and it’s still possible to login appUser on SSMS. If you are a really evil user, you can change the application name under the “Additional Connection Parameters” on connection Options.
Furthermore, even if you deny connections on SSMS you might want to keep an eye on things like SQLCMD.EXE. Nevertheless, if all you need is love a soft remainder that using the applicational user is not cool, that might do the trick. Otherwise, sorry for wasting your time. But there is still hope.
The good news is that we can use alternative approaches, if blacklisting application names is not an option. For example, we can restrict logins to a range of authorized IPs, as suggested by Solomon Rutzky in this post.
--(...)
IF (ORIGINAL_LOGIN() = N'appUser' AND
(
CONVERT(VARCHAR(10), CONNECTIONPROPERTY('net_transport')) <> 'TCP' OR
CONVERT(VARCHAR(10), CONNECTIONPROPERTY('client_net_address')) <> '10.10.1.1'
)
)
BEGIN
RAISERROR('Call the Federales',16,1)
ROLLBACK;
END;
This method implies less security risks as it would be necessary to be logged onto the production server, or spoof its IP, in order to overcome this restriction.
Finally, if you need to remove the trigger, just execute the following command. Be sure to specify “ON ALL SERVER” if the same was specified upon creation, otherwise the trigger won’t be removed.
I’m in a SQL Server 2016 database trying to view the Top Resource Consuming Queries. Since this database has Query Store enabled, I went to Query Store > Top Resource Consuming Queries and the usual screen pops-up, as the data is loading:
After a few seconds waiting, I gave up on Query Store and went to see random videos on the internet do other very professional task, that took me about ten minutes. When I switched back to check the result, the query responsible to retrieve this information was still running.
At this stage, I was wondering that maybe the Query Store got too fat to be fast. The results shown that it was occupying about 1.5 GB. That’s not very impressive, isn’t it?! But that’s something, I guess.
Initially, I thought using DBCC CLONEDATABASE. This command creates a schema-only copy of a database, to the same server as the source database and allows us to include Query Store data. After that, it would be safe to clear the Query Store without losing those statistics, and it would make a fresh start, collecting enough data to present the recent results, without the weight of historical data. Nevertheless, I would have to go through the same pain to analyze older data, so I decided to leave this path and adapt the query that the Query Store was running instead:
SELECT TOP 25
q.query_id,
OBJECT_NAME(q.object_id),
rs.avg_duration,
rs.avg_cpu_time,
rs.avg_logical_io_reads,
rs.avg_logical_io_writes,
rs.last_execution_time
FROM
sys.query_store_query AS q
JOIN
sys.query_store_plan AS p ON q.query_id = p.query_id
JOIN
sys.query_store_runtime_stats AS rs ON p.plan_id = rs.plan_id
WHERE
q.object_id > 0
AND rs.last_execution_time > DATEADD(HOUR, 1, GETUTCDATE())
ORDER BY rs.avg_duration DESC
By including Live Query Statistics, I was able to see that something was going off with the amount of data on the TVF used by the Query Store, that was posteriorly going through a Filter operator, a Row Count Spool and ended up being joined using Nested Loops.
As quick workaround, I hinted the query with OPTION(HASH JOIN) to change the plan and…
It ran in about 2.5 seconds! As we can see, the plan changed completely – No nasty Filters, no Spools, nor Nested Loops, obviously.
We can use sp_create_plan_guide to apply this principle in the query running behind the Top Resource Consuming Queries in the Query Store GUI, but removing the power of choice from SQL Server is dangerous and often ends up in tears. So, proceed carefully!
Statistics shape the way executions plans are built and have a massive impact on query performance. In order to achieve a better understanding on this relation, we will start to briefly explore what happens when a query is executed and where do statistics come into play.
Query Execution Phases
When a query is executed, there’s quite a lot happening in the Relational Engine inside SQL Server. In a higher plan, this process can be divided into four sequential phases: Parsing, Binding, Optimization and Execution. Each phase requires an input that is consumed and transformed into an output that is forwarded to the next phase.
Parsing
In the first step, a syntactic validation of the T-SQL is done, also known as query parsing. The example provided in the image below, represents a query that failed to be parsed, in that case, because the SELECT clause was miswritten.
The output of this phase is a parse tree, which comprehends the logical steps required to perform the query.
Binding
When the query is successfully parsed, query binding takes place. This process, owned by the algebrizer will resolve the parse tree, trying to bind the respective strings to specific objects within the database. This translates to, for instance, verifying if referenced tables and columns exist under the current execution context. It’s important to mention that the query can use names that do not match directly to any database object, e.g. if aliases are being used. So, it’s also the algebrizer job to resolve these cases. Additionally, it will also identify datatypes and process aggregate functions to some extent (known as aggregate binding). The following example shows a query that could not be resolved, because it refers to a column that does not exist.
What do you mean, “invalid”? How am I supposed to fix slow queries?
The output of this phase is a query processor tree, that will be handed to the query optimizer.
Optimization
Well, in fact, the query processor tree isn’t the only input received by the optimizer. Building a new plan from scratch consumes resources, specifically time and CPU. In order to manage effectively this resources, SQL Server will keep plans in cache, whenever possible, in order to reuse them. So, the optimizer will also receive the hash code of the query currently being executed, as an input. This hash code will be used to probe the existence of a valid execution plan in cache, that matches the query. If there is a match with a valid plan, the optimization process stops and the cached plan is reused. However, if changes were made to any table referenced in the query, or statistics were updated in the meantime, the plan is considered invalid and must be rebuilt.
If no valid plan is found, the optimizer creates a new plan, which in essence is an analysis of many alternate ways to achieve the expectable result. The optimizer estimates a cost for each alternative approach and tries to find a plan, cheap enough, in an incredible short amount of time. Trivial plans (e.g. SELECT Col1 FROM Table1) don’t go through this optimization process, as they typically don’t have alternative approaches to be considered, due to their simplicity. More complex queries will be subject to a full cost-based optimization process, resulting in a cost-based plan, calculated according to the cardinality estimations. After the cheapest possible plan is chosen and built, it is handed to the Execution Engine.
Execution
Finally, in the execution phase, the Execution Engine will receive and process the execution plan. Validations on the plan are reinforced at runtime, to confirm that it is still valid. If something happened in the meanwhile that changed the initial state, for example because statistics were updated, execution is paused, the compilation process is invoked and a new plan is built for the affected statement(s).
The Importance of Statistics
Statistics provide information about the distribution and selectivity of values across columns, using a sample that will represent the entire universe of values. The selectivity of a value is the ratio between the rows that pass the selection criteria and the total number of rows. We can extrapolate this principle and think in the selectivity of a column as the ratio between the number of distinct values and the total number of rows. This basically defines the uniqueness of what you are trying to find and it will provide guidance to the optimizer on how to shape the plan. In a side note, keep in mind that statistics can be manually created, (whether directly or through the creation of indexes) or automatically created by the SQL Server when a query is executed, for the optimizer to consume them.
This process is fundamental because we want our queries to run as fast as possible, what wouldn’t happen if the optimizer had to scan all data in all referenced tables.
Going Under the Hood
For the purposes of this demo, we’ll only cover the usage of single column statistics, equality predicates and single table queries. Bear in mind that the behavior differs when using multiple column statistics, joining other tables or applying more complex predicates.
First, we’ll run the following query, which will retrieve the matches where Manchester City (Id=77) played at Etihad Stadium.
SELECT * FROM Matches WHERE HomeTeam = 77
In the execution plan, we’ll see that SQL Server estimated, correctly, that 390 would be retrieved, from a total of 535 983 rows.
If we open the XML execution plan, we’ll be able to confirm which statistics were used in order to make the estimate, under “OptimizerStatsUsage“.
Next, we use DBCC SHOW_STATISTICS to view to view those statistics, related to HomeTeam column.
The information that makes up statistics is divided in three sections:
Header: Metadata about the statistics.
Density Vector: Selectivity of the data, used to measure cross-column correlation.
Histogram: Distribution of values in the first key column of the statistics object.
A histogram is composed by the following columns:
RANGE_HI_KEY: Upper-bound value for a step (always based on leftmost column).
RANGE_ROWS: Number of rows with a value falling within a histogram step, excluding the upper bound.
EQ_ROWS: Number of rows whose value equals the RANGE_HI_KEY.
DISTINCT_RANGE_ROWS: Number of rows with a distinct column value within a histogram step, excluding the upper bound.
AVG_RANGE_ROWS: Average number of rows with duplicate column values within a histogram step, excluding the upper bound calculation (RANGE_ROWS / DISTINCT_RANGE_ROWS)
In this case, because 77 hits a RANGE_HI_KEY in the histogram, we’ll be able to notice that the estimated value is on the column “EQ_ROWS” of that row.
The formulas to calculate the estimated number of rows and selectivity, can be expressed as it follows:
Selectivity is the key to combine statistics. Next, we’ll combine two single column statistics, by adding another equality predicate do the WHERE clause of the previous query. It’s important to stress that the version of the Cardinality Estimator (CE) dictates how calculations are made. The legacy CE (prior to SQL Server 2014) assumes that there is no correlation between multiple columns. Therefore, the cardinality is calculated with the following formula, that comprehends these variables: estimated number of rows (E), selectivity of the first (S1) and second (S2) columns and the total number of rows (R).
Contrarily, the new CE assumes some correlation between columns and estimates the number of rows as it follows:
Now, let’s run a new query that retrieves the games where Manchester City played at home and the full-time result was a win to the away team.
SELECT * FROM Matches WHERE HomeTeam = 77 AND FTR = 'A'
We already know the values related to the predicate HomeTeam = 77, so let’s focus on the full-time result column (FTR) and the specific value “A”. If we open the XML execution plan, we’ll notice that the statistics of another column were added.
We then repeat the DBCC SHOW_STATISTICS command, changing it to match the above highlighted stats, related to FTR column.
These statistics will show us the same 535 983 total rows and 9597 occurences of “A” in the FTR column.
Ok, now we have all the data we need, so let’s calculate the estimated number of rows for our query.
If we open the execution plan, we’ll be able to confirm that the our calculations match the ones made by SQL Server.
And as expected, if we change to the legacy CE, we will observe that the estimate has changed.
Accordingly, if we apply the formula used by the legacy CE, our results will also match:
In this case, the new CE did a better job estimating the workload, nevertheless, it doesn’t mean that the legacy CE is ready for retirement. In fact, in many cases the tables will turn and grandpa CE will do a better job. But, generally speaking, you won’t solve that many problems just by changing the CE version, despite the fact that it can have an observable impact, depending on the query specificities.
Conclusion
We’ve briefly looked into the phases of query execution and stated that the optimizer will choose the lowest-cost plan, based on its estimated cost that largely depends on the cardinality estimations. This implies that the quality of the plan choice is highly correlated to the quality of the statistics used by the optimizer. In a future post we’ll see in practice the negative impacts of bad estimates and what can be done to avoid some tears.
Implicit conversion… Sounds very clever, so it can’t be that bad, right? Well, in fact, implicit conversions can be huge performance killers and silent enough to stay off the radar, if we are not aware. The good news is that they tend to be very easy to solve, so let’s make an easy buck.
The story begins with mismatched data types on a JOIN operator or a WHERE clause that force SQL Server to implicitly convert values, on the fly. This can be as subtle as mistaking a VARCHAR for an NVARCHAR, although the consequences can be pretty awful. But, before diving into the problem itself, let’s talk about conversions and data type precedence.
Implicit vs Explicit Conversion
Data types can be either implicitly or explicitly converted.
Implicit conversions are nearly invisible to the user because they are automatically executed, i.e. they do not require the explicit use of any function.
Contrarily, in explicit conversions, CAST or CONVERT functions must be used. The recommendation is to use CAST function if ISO compliance is desirable and CONVERT function in order to take advantage of the style functionality.
Nonetheless, it’s not possible to convert every possible data type combination. The following matrix (provided by Microsoft here) shows all data type conversions for SQL Server system-supplied data types.
It’s important to notice that when an explicit conversion is done, the resulting data type is determined by the statement itself.
Implicit conversions like assignment statements, result in the data type that was defined by the variable declaration or column definition. For comparison operators or other expressions, the resulting data type depends on the rule of data type precedence.
Data Type Precedence
This rule simply states that, when an operator combines expressions of different data types, the data type with the lower precedence is converted to the data type with the higher precedence. The order is defined as follows:
1. user-defined data types (highest)
11. real
21. text
2. sql_variant
12. decimal
22. image
3. XML
13. money
23. timestamp
4. datetimeoffset
14. smallmoney
24. uniqueidentifier
5. datetime2
15. bigint
25. nvarchar
6. datetime
16. int
26. nchar
7. smalldatetime
17. smallint
27. varchar
8. date
18. tinyint
28. char
9. time
19. bit
29. varbinary
10. float
20. ntext
30. binary (lowest)
Data type precedence (Transact-SQL)
How to Uncover Implicit Conversions
There are a couple of ways to uncover implicit conversions. The most usual is, probably, by analyzing the Execution Plan. If we’re inspecting the execution of a query where an implicit conversion that may affect cardinality estimate happened, a warning will be displayed. We’ll further look into this.
It’s also possible to capture these occurrences through the use of DMVs. I’ve seen some examples using sys.dm_exec_query_plan DMV in order to accomplish it, but we’re going to use the brand new and awesome sys.dm_exec_query_plan_stats DMV brought by SQL Server 2019. We’ll take a deeper look into it in a future post, but for now just keep in mind that this DMV will show us the last actual plan for a given query. Also, make sure to activate the scoped configuration as specified below:
ALTER DATABASE SCOPED CONFIGURATION SET LAST_QUERY_PLAN_STATS = ON;
--ImplicitConversions
DECLARE @DBName VARCHAR(20) = 'FStats';
SELECT TOP 30
t.text AS 'Query',
qs.total_worker_time/qs.execution_count AS 'Worker Time (Avg)',
qs.total_elapsed_time/qs.execution_count AS 'Elapsed Time (Avg)',
qs.total_logical_reads/qs.execution_count AS 'Logical Reads (Avg)',
qs.total_physical_reads/qs.execution_count AS 'Physical Reads (Avg)',
qs.max_elapsed_time AS 'Max Elapsed Time',
qs.max_logical_reads AS 'Max Logical Reads',
qs.max_physical_reads AS 'Max Physical Reads',
qs.execution_count AS 'Execution Count',
qp.query_plan AS 'Query Plan'
FROM
sys.dm_exec_query_stats AS qs WITH (NOLOCK)
CROSS APPLY
sys.dm_exec_sql_text(plan_handle) AS t
CROSS APPLY
sys.dm_exec_query_plan_stats(plan_handle) as qp
WHERE
t.text NOT LIKE '--ImplicitConversions%' AND
CAST(query_plan AS NVARCHAR(MAX)) LIKE ('%PlanAffectingConvert%') AND
t.[dbid] = DB_ID(@DBName)
Finally, another more comprehensive approach might be setting an Extedend Event to collect implicit conversions happening in the database. Creating a plan_affecting_convert event will allow us to uncover these CONVERT_IMPLICIT warnings. For the purposes of this demo it will be configured as it follows:
CREATE EVENT SESSION [xe_ImplicitConversions] ON SERVER
ADD EVENT sqlserver.plan_affecting_convert
(
ACTION(sqlserver.transaction_id)
WHERE ([sqlserver].[database_name]=N'FStats')
)
WITH (STARTUP_STATE=ON)
Implicit Conversions Woes
But why should we care about implicit conversions? Does it impact performance that much? Let’s find out! The following query retrieves the disciplinary sheet of football matches where 4 or more players were sent off.
DECLARE @NumRedCards tinyint = 4;
SELECT
MatchDetails,
HF HomeFouls, AF AwayFouls,
HY HomeYellowCards, AY AwayYellowCards,
HR HomeRedCards, AR AwayRedCards, TRC TotalRedCards
FROM
MatchesDetails
WHERE
TRC >= @NumRedCards
Everything looks fine, right? Despite that, if you sense evil lurking, you’re absolutely correct. Using the previous xe_ImplicitConversions session, we’ll be able to see this:
Accordingly, the set of DMVs specified above, will reveal conversion problems that may affect cardinality estimate in the query plan choice.
If we click on the query plan and then hover on both operators, the following information will pop up:
Besides the obvious warnings, it’s clear that a nasty Clustered Index Scan was done and it’s quite strange that SQL Server didn’t cry for an index, isn’t it? Let’s create the bloody thing and ignore the implicit conversion, for now.
CREATE NONCLUSTERED INDEX IDX_MatchesDetails_TRC_Includes ON MatchesDetails(TRC)
INCLUDE(MatchDetails,HF,AF,HY,AY,AR,HR)
WITH(DATA_COMPRESSION=PAGE)
Noice! Now, with an unstoppable confidence, we will rerun the query and the plan will unequivocally show us that we’ve…
Failed. Miserably. An ugly Clustered Index Scan is still happening and the logical reads remain at 161 516.
Ok, so let’s change the approach here. First by dropping the index:
DROP INDEX IDX_MatchesDetails_TRC_Includes ON MatchesDetails
Then by analyzing the warning presented in the execution plan. According to it, there’s a data type mismatch between TRC and @NumRedCards. If we look into the definition of MatchesDetails table, we’ll confirm that TRC is, in fact, a varchar – genius.
Solving the Case
Elementary, my dear Watson.
Although the correct path here would be to alter the column data type, for simplicity’s sake we’ll just change the data type of @NumRedCards to varchar(2) and rerun the query.
DECLARE @NumRedCards varchar(2) = 4;
SELECT
MatchDetails,
HF HomeFouls, AF AwayFouls,
HY HomeYellowCards, AY AwayYellowCards,
HR HomeRedCards, AR AwayRedCards, TRC TotalRedCards
FROM
MatchesDetails
WHERE
TRC >= @NumRedCards
This time, although the Clustered Index Scan hasn’t vanished, we got a missing index hint that corresponds exactly to the one created earlier. After recreating it, and rerunning the last query, the result will be the following:
Success! Just like a rainy cloud, the warning has disappeared and the day is sunny again. We’re now efficiently using the index with an Index Seek and the logical reads dropped to a glorious 4.
Cry Me a River
There’s nothing much left to say, in this context, regarding how implicit conversions can mess up the cardinality estimates and consequently ruin SARGability, index usage, and burn CPU. Despite that, it’s relevant to mention that implicit conversions aren’t always a problem, even if SQL Server says so. One example of that is when the conversion occurs on the SELECT statement, after the rows are fetched. SQL Server will have an embarrassing outbreak, threatening you with the same old warnings, stating that cardinality estimate can be affected, but it’s all a lie, it certainly won’t. Nevertheless, stay alert for the real performance killers! 🖖
As suggested by the title, I like redundancy, bad titles and we’re going to have a brief look into MemoryGrantInfo.
In a previous post, we’ve talked about an awesome feature from the new SQL Server 2019 – Row Mode Adaptive Memory Grant Feedback. We saw this feature in action and briefly explored MemoryGrantInfo, but some fields were not explained. So, let’s recall the MemoryGrantInfo section and analyze its properties:
DesiredMemory: Amount of memory (in kilobytes) estimated to be needed to run the execution plan, based on the cardinality and data size estimation. This value will be cached as part of the plan and will remain the same – if Adaptive Memory Grant Feedback is not enabled.
GrantedMemory: Amount of memory that was actually granted at runtime. This value can be lower than DesiredMemory if the server doesn’t have enough available memory at the moment of the grant. Due to this fact, it’s possible that its value differs across similar executions, depending on the server’s available workspace memory at the time. When the grant isn’t high enough to keep all data in memory it will write it into tempdb – this is also known as a spill and it can degrade query performance. This doesn’t mean that a lower grant will always cause a spill, nor that a spill will always have a substantial impact on performance, nevertheless, that’s definitely something to keep an eye on.
GrantWaitTime: Time (in seconds) that the query had to wait before it was granted the memory to execute. This will translate into RESOURCE_SEMAPHORE waits.
IsMemoryGrantFeedbackAdjusted: Current state of Adaptive Memory Grant Feedback, that will be represented by one of the following values: NoFirstExecution, YesAdjusting, YesStable, NoAccurateGrant, NoFeedbackDisabled. For more details, check this post about Adaptive Memory Grant Feedback.
LastRequestedMemory: Indicates RequestedMemory value from the previous execution. While trying to calculate the ideal memory grant, Adaptive Memory Grant Feedback will use the requested value from an execution in order to improve the following. This property will allow us to easily check how the grant is being adapted.
MaxQueryMemory: Maximum memory allowed for a single query, according to Resource Governor’s MAX_MEMORY_PERCENT configuration, which by default is 25% of total query memory. In a case where operators are spilling data into tempdb and estimates are fairly correct, we might be experiencing memory starvation.
MaxUsedMemory: Amount of memory actually used by the query during execution.
RequestedMemory: Amount of memory requested, based on RequiredMemory, DesiredMemory, SerialRequiredMemory, SerialDesiredMemory, and MaxQueryMemory. When any query is executed, SQL Server Engine checks if the desired memory exceeds the Max Query Memory or not, if it exceeds, then it reduces the requested memory by adjusting parameters as Degree Of Parallelism (DOP), until it fits in.
RequiredMemory: This property represents the required memory to create the respective internal data structures, for a given DOP, when the query runs in parallel. If the query is running in serial mode, RequiredMemory value will be the same as SerialRequiredMemory. It’s mandatory that this amount of memory is available, in order to the query to begin execution.
SerialDesiredMemory: Amount of memory estimated to be needed to run the execution plan on serial mode, without spilling to disk.
SerialRequiredMemory: Required memory for a serial query plan to execute. It’s mandatory that this amount of memory is available in order to the query to begin execution.
