An SQL text by Erland
Sommarskog, SQL Server MVP. Most recently updated 2018-06-15.
Copyright applies to this text. See here for font conventions used in this article.
In this article I discuss various checks that SQL Server could perform when you create an SQL module (that is, a stored procedure, trigger etc) that would help you to find stupid errors early. By finding the errors earlier you would be more productive – and your testers will never have to deal with your silly goofs.
Since many of these checks would break existing code if they were introduced right off, the idea is that these checks are only active if a certain setting is in force; tentatively the command would be SET STRICT_CHECKS ON. In March 2007, I submitted a suggestion for this feature on the old Connect site and today you find it on the modern Uservoice site as https://feedback.azure.com/forums/908035-sql-server/suggestions/32890552-add-optional-checks-for-more-robust-development.
The purpose of this article is to develop this idea further with the hope that it will serve as inspiration for the SQL Server team, as they work on further enhancements to SQL Server. (It hasn't happnened yet, but the optimist in me never gives up!)
SET STRICT_CHECKS ON would be a compile-time setting. When a batch is entered and parsed, a number of things that goes unnoticed today would yield a compilation error or possibly a warning. This applies to CREATE PROCEDURE, CREATE FUNCTION etc as well as ad-hoc batches.
The first section is a discussion on general principles, but the main body of this article is devoted to the possible checks that could be performed when SET STRICT_CHECKS ON is in force.
Table of Contents
At first glance, the whole idea with optional checks may seem corny, but there are precursors. Two
other languages that I have worked with have similar features. Visual Basic has
Option Explicit and Perl has
use strict. And people with
experience of C from older days may recall the lint program.
The purpose with these checks is to help the programmer to find silly typos and goofs early, so he don't have to spend his time to proof-read the code for mistakes that the machine easily can detect. Or avoid egg on his face when his typo halts testing or even worse production.
In this document I assume that the command to enable this feature would be SET STRICT_CHECKS ON, and I refer to it "strict checks in force" etc. The SQL Server team may prefer something else, like SET STRONG_CHECKING ON and I trust their good judgement in that regard. I have gone for strict checks consistently through this document, as I also use it in the feedback item on Uservoice.
Microsoft may prefer something different than a SET option. Two alternatives that come to mind are:
In this document, I assume that it is a SET option, but that is only to keep the discussion simple. In fact, when I discussed this with a contact at Microsoft he said that if they were to do this, he preferred it to be a database setting rather than a SET option.
This document suggests a wide range of checks, of which some are quite aggressive. This raises the question whether there should be a single setting or a couple so that you can opt out of some checks. This is not without a precursor: in Perl,
strict has three options, so that you can say
use strict qw(vars
subs) if you only want those two checks. Thus, in SQL Server we could
have SET STRICT_CHECKS OBJECTS ON, SET STRICT_CHECKS IMPLICIT_CONVERSION ON etc,
and of course also a SET STRICT_CHECKS ALL ON.
However, this would increase the testing matrix for Microsoft. At the same time, it could contribute to make the feature more difficult to use: Surely, best practice would mandate SET STRICT_CHECKS ALL ON, so if only some checks are in effect that would be confusing. Thus my preference and suggestion is that there should be a single all-or-nothing knob for strict checks, and this is what I assume in the rest of this document.
My list of possible checks is tentative, and I more or less expect the SQL Server team to discard some of them. There is one important thing to note though: Often when you introduce a new feature, some whistles and bells are deferred to later versions. Usually, this is a good idea, but for this feature this could be problematic. Many of the new checks will break existing code, and for the first version that is alright, because you need to say SET STRICT_CHECKS ON for it to happen. But once the setting has shipped, Microsoft cannot add new checks in the future versions of SQL Server without breaking backwards compatibility for applications that have embraced strict checks v1.0 – butg which may not be ready for v2.0.
So realistically, the first implementation of this feature will also have to be the complete feature. In later versions, there can only be new differences between strict and lax checks for new features. I can see minor changes being covered by compatibility levels, but not strict checks for an entirely new area.
This is quite a challenge, but keep in mind that this applies to the box product only. In Azure they could make parts available in a preview where they have no obligations about backwards compatibility. And in this way the feature can evolve with user input.
Should the issues that strict checking reveals be raised as errors or as warnings?
The advantage of letting them be only warnings is quite obvious: there would not be any real compatibility issues. Taken to the extreme, there would not have to be any SET command at all, but the checks could always be in force.
However, there are also a number of disadvantages. One is how errors and warnings work with the tools today. With errors, SQL Server reports a line number, and this line number is displayed in the query tool, and the programmer can double-click on the line to quickly find the error.
For warnings on the other hand, SQL Server does not seem report a correct line number, and the query tools do not display them. This means that while the programmer is informed of the problem, it will take him longer time to find where the problem is.
This particular issue could be addressed though: SQL Server could report all strict-check issues as level 9, and make sure to report the correct line number for the issues. The tools would then be changed to display procedure and line number for level-9 messages. Alternatively, the messages could be reported as level-0 messages as warnings are today, but all strict-check messages would start with "Line nnn:". The tools would need to be adapted so that you can double-click on such a message to find where it origins from.
But there is more to it. I am quite sure that once these checks are in place more than one DBA would say "I don't accept any strict-check messages in my databases", and he will want to have them all reported as errors to prevent the objects to be created. This could be met by having SET STRICT_CHECKS WARNINGS, SET STRICT_CHECKS ERRORS and SET STRICT_CHECKS NONE. But again, my preference is for a simple on/off switch.
What is a little more realistic is that only some of the issues found by strict checks are reported as errors, whereas others are reported as warnings. And that is alright, as long as line numbers are reported for these warnings.
Nevertheless, to simplify this text, I assume that all issues found by strict checks are reported as errors and I don't discuss the possibility of reporting them as mere warnings any further.
One more thing: all errors detected by strict checks should make it clear that they come from strict checks and possibly inform the user how to turn off these checks.
The purpose of SET STRICT_CHECKS ON is to control compile-time behaviour. Therefore, it raises some questions would it mean if SET STRICT_CHECKS ON (or OFF) would appear in the middle of a stored procedure. Consider:
IF @x >= 0 SET STRICT_CHECKS OFF ELSE SET STRICT_CHECKS ON
Since SET STRICT_CHECKS is a compile directive, what would this mean?
To avoid all such ugly problems, my suggestion is that the command is only permitted in the top-level scope, and it must either be in a batch of its own, or be the first command in a batch (in the latter case, it would serve as a compile-time directive). And when I say the first command in a batch, I mean it. That is:
CREATE PROCEDURE some_sp @var int = 99 AS SET STRICT_CHECKS ON
will not fly. Some of the checks that I propose affects the procedure header, so above the raises the question, is the check in force for the header or not?
Strict checks are there to help the programmer to catch typos and goofs. But depending on how the checks are implemented, there may be situations where the checks gets in his way, because he actually wants to do what he types. In a few places in this document, I have identified situations where this could occur.
Since turning off strict checks for an entire stored procedure would be to throw out that proverbial baby with the equally proverbial bathtub, I thought about alternatives. One alternative would be to have BEGIN NOSTRICT and END NOSTRICT and within this block strict checks would be turned off. However, this would invite to bad practice, where inexperienced programmers would enclose all their procedures in this block, because they don't like the error messages.
In this document I have instead settled for a setting that works on line level. If the programmer adds the comment
/* NOSTRICT */ on a line, SQL Server will not report any strict-check errors on that line. This is akin to how the old lint program worked.
Admittedly, it would be best to be without a general loophole to keep the language clean. But that would also require that there are ways out when you have legit reasons to work against spirit of the rules. Sometimes such ways out are easy to identify. E.g., I suggest that it would be illegal to assign a varchar(10) variable to a varchar(5) column, since this could cause data loss. The way to work around this is to use cast or convert. But there may be situations where Microsoft would have to introduce new language constructs to avoid unwanted roadblocks. Since I did not want not wander into such territory, I have opted for general loophole with
/* NOSTRICT */ in this document.
Strict checks are intended to help the developer, but SQL Server does not know
if there is a developer or an application on the other end. If an application
issues SET STRICT_CHECKS ON, and then runs ad-hoc batches, they would be
checked. I don't see any major problems with this. Stupid errors like
tbl a ON a.col = a.col would result in an error, which is probably better
than returning an incorrect result.
A more intriguing situation is when SQL Server compiles an existing stored procedure to build a query plan. Would strict checks apply in this case? I can see some advantages with this. Today, if a procedure refers to a non-existing table, it bombs when you reach that statement, in which case the procedure is terminated, even if there is a local catch handler. This can cause some mess if the procedure has started a transaction. With strict checks in force, the compilation would fail directly and the behaviour would be cleaner. But I don't see this as mandatory. If Microsoft finds it easier to compile code already stored in SQL Server in unstrict mode, I think this would be alright..
Since SET STRICT_CHECKS ON is a compile-time setting, should it be saved with the SQL module, and in such case what would that mean?
Today there are two settings which are saved with the module: ANSI_NULLS and QUOTED_IDENTIFIER, and when you run the procedure, the saved settings apply. This has caused quite some confusion over the years, and it is definitely not desirable. So if the setting is saved with the procedure, it would be informational only: to make it possible for the DBA to review whether there are any procedures in the database that were entered with strict checks off.
However, observe that even that if even if a procedure was entered with strict checks in effect, that is no guarantee that it will execute or compile successfully at run-time, since a dropped table or column may have invalidated the procedure. Thus, saving the setting in metadata is at best a nice-to-have feature.
Since strict checks is a engine feature, the impact on the tools is small. There are a few points, though:
->Advanced, there should be a check box for SET STRICT_CHECKS ON, so that you can always get this setting when you connect. For the first release of this feature, this checkbox should probably not be checked by default, but if the feature is well received this could be changed for the following version.
When I said above that nothing has happened since I first wrote this article, that was not 100 % correct. With SQL Server 2012, Microsoft released SQL Server Data Tools (SSDT) and SSDT performs some of the checks I suggest, although far from all. The checks are performed outside SQL Server, but they use a language service which, as I understand, uses the same parser and binder as the engine.
SSDT is definitely a commendable effort. I've played with it a bit, and my overall impression is positive. At the same time, SSDT is a very complex tool and introduces a way or working for development and deployment that your organisation may not be ready for. That is, SSDT is not for everyone. And therefore SSDT is not a solution for the proposals in this article. It's a decent workaround for some, but in the long run, this should be in the engine.
In the following text, I point out a few checks that I have noticed are available in SSDT. There could be others that I have not noticed; I have not played that extensively with SSDT.
I'm a long-timer with SQL Server and I first came in touch with the product in 1991. Back in those days, if you said something like:
CREATE PROCEDURE bad_sp AS PRINT 'This prints' SELECT col FROM nonexisting
You would get the error message:
Server: Msg 208, Level 16, State 1, Procedure bad_sp, Line 3
Invalid object name 'nonexisting'.
already when you tried to create the procedure. This was the state of affairs up to SQL 6.5. With SQL 7, Microsoft introduced what they call Deferred Name Resolution. You no longer get an error about the missing table, but the procedure creates just fine. When you run it, you get this output:
Msg 208, Level 16, State 1, Procedure bad_sp, Line 3
Invalid object name 'nonexisting'.
That is, it is not until the SELECT statement actually executes that you get the error message. When I first heard about this, I could not just believe it! I was using SQL Server to develop a large enterprise system, and Microsoft changes the behaviour as if SQL Server was only for toys. Imagine an important function grinding to a standstill just because of a small typo that SQL Server could have caught up front! You may object that such typos should be caught in testing, and in most cases they do, but: 1) the typo may be in an odd code path that was not covered by the testers, 2) if the typo stops the tests, the testers will have to wait for the next build, and the company loses time and money.
There are also some other nasty side effects. Let's say that you have an outer stored procedure that calls an inner stored procedure that first starts a transaction, and then later runs into an error with a missing table. The inner procedure is aborted, but execution continues in the outer procedure – with the transaction still active! How many programmers are prepared for that? TRY-CATCH makes this is a little easier to handle, but if the inner procedure has a TRY-CATCH, that CATCH will not fire, only the TRY-CATCH in the outer procedure.
Another problem with deferred name resolution is that the missing table can mask other errors. Consider this procedure:
CREATE PROCEDURE some_sp @productid int AS CREATE TABLE #temp (orderid int NOT NULL, orderdate datetime NOT NULL) SELECT o.OrderID, o.OrderDate FROM Orders WHERE EXISTS (SELECT * FROM #temp od WHERE od.orderid = o.OrderID)
If you look closely, you see that the programmer has failed to specify the alias he is using for the Orders table. Nevertheless the procedure is created without any objections. But if you run it, you will get a slew of error messages:
Msg 4104, Level 16, State 1, Procedure some_sp, Line 4
The multi-part identifier "o.OrderID" could not be bound.
Msg 4104, Level 16, State 1, Procedure some_sp, Line 4
The multi-part identifier "o.OrderID" could not be bound.
Msg 4104, Level 16, State 1, Procedure some_sp, Line 4
The multi-part identifier "o.OrderDate" could not be bound.
In this examples, I used a left-out alias definition as an example, just to emphasise how bizarre this is. You get exactly the same result with the alias in place, but for instance misspell OrderDate as OrderDte. That is, common sloppy errors you want the computer to find for you go unnoticed.
When SET STRICT_CHECKS ON is in effect, a statement must only refer to known objects and to known columns in those objects. Plain and simple.
The reader may object that strict checks is no full guarantee that code will not blow up at run-time because of missing tables, since a table could be dropped after the procedure has been created. This is true, but the intention of strict checks is not to make SQL Server fool-proof; it is to help the programmer to catch silly errors early. In my experience it is not that common that tables are accidently dropped. Silly typos are far more common. At least if I am the one who does the typing.
What about dynamic SQL? Of course, if your stored procedure creates dynamic SQL, strict checks are not going to help you to catch those errors before run-time. But in an enterprise system, most of the code should be in stored procedures with static SQL. At least in my not-so-humble opinion.
Note: a possible alternative here is the clause WITH SCHEMABINDING which exists already for functions and views, and it also required natively compiled modules. This clause in the procedure header requires that all objects referred to in the module do exist. Furthermore, you cannot drop an object which is referred by a module WITH SCHEMABINDING. In my opinion, extending this to traditional stored procedures is not going help what I'm aiming for here. It may work with natively compiled modules since you typically use them in a small corner of your database. But in a system with 1700 tables and 8000 stored procedures (yes, I work with such a system), WITH SCHEMABINING will be too rigid to be workable.
In the following sections, I will closer at what strict checks implies for different object types.
The biggest challenge with designing strict checks is how to deal with temp tables, and I will discuss two different approaches.
We saw above that the errors we got above was due to a query that referred to a temp table that was defined within the procedure. And if you started using SQL Server on SQL 2000 or later, you may find it all natural. When the procedure is created, the temp does not exist, so how could SQL Server do any better? Obviously we don't want any error or even a warning for this missing temp table!
But recall what I said: deferred name resolution was introduced in SQL 7. Yet, temp tables created in the procedure did not result in any error in earlier versions. For instance, this procedure creates and executes without an error on SQL 6.5 as well:
CREATE PROCEDURE good_sp AS CREATE TABLE #tmp(a int NOT NULL) SELECT a FROM #tmp
But change the procedure a little bit:
CREATE PROCEDURE another_bad_sp AS CREATE TABLE #tmp(a int NOT NULL) SELECT b FROM #tmp
On SQL 7 and later, this procedure is created without an error, and then bombs at run-time because of the missing column. However, if I try to create it SQL Server 6.5, I am told:
Server: Msg 207, Level 16, State 2, Procedure another_bad_sp, Line 3
Invalid column name 'b'.
If SQL 4.2 and 6.5 could do it, why not later versions?
What SQL 6.5 and earlier versions did was to read the procedure code and extract all temp table defintions, be that through CREATE TABLE or SELECT INTO and then derive their schema. Thus, with strict checks in force, modern versions of SQL Server would do the same.
Now, since CREATE TABLE is a run-time statement, this raises the question about what should happen if the CREATE TABLE statement is within an IF statement or you have something as stupid as:
CREATE PROCEDURE another_bad2_sp AS SELECT a FROM #tmp CREATE TABLE #tmp(a int NOT NULL)
While it's relatively simple to find this particular error, flow analysis gets hairy when you add control-of-flow statements into the mix. So I find it perfectly acceptable if SQL Server would let this pass without an error. (But SQL 6.5 realises that this is wrong and refuses to create the table.)
You may ask: what if we want to have different definitions of his temp table, like this::
CREATE PROCEDURE doubletmp @i int AS IF @i = 0 CREATE TABLE #tmp(a int NOT NULL) ELSE CREATE TABLE #tmp(b int NOT NULL)
This is the least of worries, because here is something amazing: all versions of SQL Server from 6.5 and up produce the same (somewhat inaccurate) error message when you try to create it:
Msg 2714, Level 16, State 1, Procedure doubletmp, Line 5
There is already an object named '#tmp' in the database.
So, SQL 7 and later do notice that there is a temp table being created in the procedure. They just don't care to use that knowledge when checking other queries.
What if the temp table exists when procedure is created? That is, you have:
CREATE TABLE #tmp(col_a int NOT NULL) INSERT #tmp (col_a) values (12) go CREATE PROCEDURE another_sp AS CREATE TABLE #tmp(col_a int NOT NULL) SELECT col_a FROM #tmp
In SQL 6.5 you get an error when you try to create this procedure, but I don't think this is the right behaviour. The CREATE TABLE command in the procedure should take precedence. The temp table that exists now is not likely to exist at run-time.
I have here confined the discussions to temp tables, since this issue is about temp tables in 99 % of the time. But if the procedure creates a static table, it should be handled in the same way. That is, SQL Server should extract the definition, and use the definition when checking the queries with one difference to temp tables: if the table already exists, this should be considered an error.
So far this path seems quite simple. It gets more difficult in the case you want create a temp table in one procedure and read or write it in a different procedure, something like this:
CREATE PROCEDURE outer_sp AS CREATE TABLE #tmp(...) EXEC inner_sp SELECT ... FROM #tmp WHERE ... go CREATE PROCEDURE inner_sp AS INSERT #tmp (...) SELECT .... go
SQL 6.5 did not have a good story here. The file that created inner_sp had to read something like this:
CREATE TABLE #tmp(...) go CREATE PROCEDURE inner_sp AS INSERT #tmp (...) SELECT ...
That is, you had to put copy of the definition of #tmp in the file, which meant that you had to have the definition for the temp table in two places, which obviously is a source for errors.
One alternative would be to use
/* NOSTRICT */ on all lines where the temp table
CREATE PROCEDURE inner_sp AS INSERT #tmp /* NOSTRICT */ (...) SELECT ...
This could be deemed acceptable, since this type of procedures is not that common. Then again, they are by no means rare, either, so there is reason to try to find something better.
Table types looks like a perfect fit here. If you could say:
CREATE TABLE #tmp AS my_table_type
The file for inner_sp could read:
CREATE TABLE #tmp AS my_table_type go CREATE PROCEDURE inner_sp AS INSERT #tmp (...) SELECT ...
And there would not really be any issue with the temp-table definition appearing in two places.
You could argue that it still would be nicer if this somehow could be stated within the procedure body. If nothing else, this would make it possibly for a tool like SSDT (see further the end of this chapter) to handle this situation. Here are two syntactic suggestions:
CREATE PROCEDURE inner_sp WITH TABLE #tmp AS my_table_type AS CREATE PROCEDURE inner_sp AS ... REFERENCES TABLE #tmp AS my_table_type
That is, in the first example it is stated in the procedure header, in the second in the procedure body. I choose WITH in the first case, since WITH is already used in the header. I chose REFERENCES in the second case, since that is already a reserved keyword, and it sort of fits.
An advantage with the first solution is that this permits for a very early error if inner_sp is called without a #tmp of the correct type existing. In fact, it may even be caught when the caller is created.
But I know of situations where the inner procedure checks if the temp table exists and creates if not. With REFERENCES this could be implemented this way:
IF object_id('tempdb..#tmp') IS NOT NULL REFERENCES TABLE #temp AS my_table_type ELSE CREATE TABLE #temp AS my_table_type
Although this may be a little bit over the top.
The reader may at this point ask what will happen if columns are added or dropped from temp table with ALTER TABLE. I guess this will require
/* NOSTRICT */.
In this section I look will at a completely different solution for the problems with temp tables, to wit one that already exists in SQL Server: table variables.
Table variables are declared objects, so if you say:
CREATE PROCEDURE some_sp @productid int AS DECLARE @temp TABLE (orderid int NOT NULL, orderdate datetime NOT NULL) SELECT o.OrderID, o.OrderDate FROM Orders WHERE EXISTS (SELECT * FROM @temp od WHERE od.orderid = o.OrderID)
This fails with the error messages we can expect, and the same is true for the other examples we looked at. No deferred name resolution here.
However, table variable are notorious for being performance busters. They don't have distribution statistics, and for this reason they don't trigger recompilation. The net effect is that the optimizer often estimates them to have one row, which can lead to disastrous plan choices.
But why cannot table variables and temp tables be married together? The most radical would be to throw away the current table variables and let
DECLARE @temp TABLE be syntactic sugar for
CREATE TABLE #temp. However, this is bound to cause performance regressions for some customers, for instance of all the recompilation that would be triggered. There are also functional aspects, as the fact that table variables are not affected by rollback while temp tables are.
But imagine something like this:
DECLARE @temp TABLE (orderid int NOT NULL, orderdate datetime NOT NULL) WITH STATISTICS
That is, we tack on an extra clause. This sort of table variable, would only be like the current table variables syntactically. But else they would be like the current temp tables, with some differences:
That is, on these points the behaviour would be the same as for regular table variables. And most importantly, compilation errors in queries with these disguised temp tables would not go unnoticed, even when strict checks are off!
When it comes to the existing syntax CREATE TABLE #tmp, you can still use it, but this type of tables would still be subject to deferred name resolution, even with strict checks in effect. That is, you would have to change your code to get benefit of this change, but since the purpose is to make development more robust, I see this as acceptable. The one situation where you would have to fall back to CREATE TABLE #tmp is when you want to add columns dynamically at later stage, a scenario I would consider rare.
At this point the reader may say: What about SELECT INTO? One solution that appears as palatable is this:
DECLARE @mytable TABLE AS (SELECT ... FROM ... WHERE ...) WITH STATISTICS
When the procedure is created, all tables in the query must exist; there is no deferred name resolution even if strict checks are off. This is required, so that SQL Server can determine the metadata for the table variable at compile time. This means that such a query cannot refer temp tables with #, but table variables declared earlier can be referred to. If the schema for a table in the SELECT query is changed, the plan for the procedure of flushed from the cache, since that could affect the schema of @mytable. At run-time, the query is excecuted when the DECLARE statement is reached. If the DECLARE statement is in a loop, it may be executed multiple times, adding more rows to the table.
The syntax above does not permit for defining indexes in @mytable. Syntax could be added for that, but it seems easier to permit CREATE INDEX for table variables WITH STATISTICS.
What about table types? Should it be possible to create a table type WITH STATISTICS? Yes, it should, because it is more or less required for the situation when you create a temp table in an outer procedure and write to it in an inner procedure. For this we would like to use table-valued parameters that are read-write. Here is an idea that permits for a relatively cheap implementation for Microsoft.
The final thing to observe about this approach is that a procedure like this would cause a compilation error when strict checks are in effect:
CREATE PROCEDURE spooky AS CREATE TABLE permanent (a int NOT NULL) INSERT permanent(a) VALUES (12)
This is because in this solution, SQL Server would never try to extract table definitions from the procedure. If you need to do this, you would have to use
/* NOSTRICT */.
I think that of the two approaches I have presented, I prefer this latter one. Although, you can certainly argue that table variables having quite different behaviour depending on the presence of WITH STATISTICS is bound to confuse people even more than the current differences between table variables and temp tables.
I will now leave the area about temp tables and cover some other situations where deferred name resolution raises its ugly head.
Say that a procedure has a query that includes an index hint for a non-existing index:
CREATE PROCEDURE bad_hint AS SELECT col FROM mytbl WITH (INDEX = no_such_index)
Today, SQL Server creates the procedure without any mention of the missing index and then the query blows up at run-time. This behaviour is clearly not acceptable. This seems like an obvious case for strict checks: if an index hint refers to a non-existing index, this is a compile-time error.
However, in this particular case, there is an alternative. There is a feedback item on Uservoice that suggests that there should not be any run-time error when the index in a hint is absent, something I entirely agree with. There is some chance that the hint itself is obsolete, and the query executes just fine anyway. Or at least no slower than the system still can progress. Whereas the now you get an run-time error which is more likely to cause an outage.
If this feedback item is implemented, should a hint for a missing index still result in an error with strict checks in force? You could argue that it may be a typo, and therefore the programmer is helped by being informed about the error. But you may also have a reporting database which is a copy of production but has extra indexes added after RESTORE, and in that case, you don't want this error to occur when you create the procedure in the production database. This could handled with a
/* NOSTRICT */ comment. But maybe it would be sufficient to issue a warning in this case.
User-defined scalar functions are also victims to deferred named resolution. For instance, SQL Server gladly creates:
CREATE PROCEDURE bad_udf @OrderID int AS SELECT dbo.no_such_udf(OrderID) FROM Orders WHERE OrderID = @OrderID
and then it bombs when you run it. This procedure should of course not be accepted when strict checks are in force.
Stored procedures are a special case, since you get a warning if you call a non-existing stored procedure:
CREATE PROCEDURE bad_call AS EXEC no_such_sp
The message for this condition has varied over the versions, but it has been there since SQL Server 4.x at least. The current wording of the message reads:
Originally the purpose of the warning was to inform the user that SQL Server was not able to add any rows to sysdepends, later sys.sql_dependencies. That particular piece of information is not that interesting, but what is interesting is of course that the procedure does not exist. SQL 2008 added a new structure for dependencies where the dependencies are stored by name, so technically there is no longer any reason for the message. And indeed in some CTP of SQL 2008, the message was gone. As you may imagine, that made me very angry. Microsoft took reason and the message is still there.
With strict checks in force the warning should be promoted to an error (because as I discussed above this makes it easier to find where this bad call is). Should the warning still be there when strict checks are off? I suggest that it should, but I promise not to make a fuzz if Microsoft removes it.
Occasionally, you may have a cross-dependency: stored procedure A calls B,
and B in its turn includes a call to A. In this case, you should probably use
/* NOSTRICT */ to get around it. (And in this case, you probably don't
want the warning at all.) An alternative is to first create A as a dummy, and then create B and alter A to have its actual contents.
You can use variables with EXEC for indirection:
In this case, there should of course not be any message at compile-time.
Service Broker objects are also subject to deferred name resolution. For instance, this
CREATE PROCEDURE sb1 AS DECLARE @dialog_handle UNIQUEIDENTIFIER; BEGIN DIALOG CONVERSATION @dialog_handle FROM SERVICE no_such_service TO SERVICE 'the_to_service' ON CONTRACT no_such_contract; SEND ON CONVERSATION @dialog_handle MESSAGE TYPE no_such_type RECEIVE * FROM no_such_queue
is accepted in SQL Server today. With strict checks, there would be errors all over the place. The only thing to note is that the TO SERVICE should not be checked. This is a string literal, and this can be a service in a remote database in a remote server so it is not possible to validate.
Would you believe it, if you say
BULK INSERT tbl FROM 'C:\temp\myfile.bcp'
the file name is actually subject to deferred name resolution. Or else, how can you explain this.
BEGIN TRY BULK INSERT tbl FROM 'C:\temp\nosuchfile.bcp' END TRY BEGIN CATCH PRINT 'This does not print' END CATCH
That is, if the file does not exist, the CATCH handler is not invoked, because the procedure is terminated on the spot (a CATCH handler in an outer scope can catch the error). Normally, this happens only with compilation errors, so I will have to assume that when SQL Server originally creates the procedure, it checks whether the data file is there, and if it's not it defers compilation until later, and if the file is still missing at run-time, this is the same as a missing table.
Now, why it would look at the data file at all when creating the procedure is beyond me. In any case, I don't think it would be a good idea if you would get an error message for a missing file even when strict checks are in force. A missing file should be handled as a pure run-time error, be that a data file or a format file, so that a local CATCH handler can work. That is what most programmers would expect anyway. (If Microsoft makes this a pure run-time check, there is also no reason that the file could not be specified through a variable, but that's another story.)
There is also OPENROWSET(BULK...), which permits you to select columns and filter rows from the data file. The column names must come from somewhere, and they can be specified in one of these two ways:
FROM OPENROWSET(BULK...) AS table_alias(column_alias,...n) FROM OPENROWSET(BULK...) AS table_alias
The first one leaves no mystery, as the column names are in the query. But in the second case the column names are taken from the format file, which is mandatory with OPENROWSET(BULK...) (unless you specify one of the SINGLE_LOB options). What would happen here when strict checks are in force? Should you get an error if you use the second form and the format file is not there? That appears impractical. It does not seem a good practice to me to rely on a source outside the database to provide column names anyway, so I suggest that the second form would be illegal when strict checks are in force.
CREATE SYNONYM mybadsyno FOR nosuchobject
should this be permitted when strict checks are enabled? I see no harm if so. As long as the synonym is not used, there is no problem. Of course, if you have
CREATE PROCEDURE bad_synonym AS SELECT col FROM mybadsyno
attempting to create that stored procedure when strict checks are in force would yield an error message.
Now, this is a funny case. Consider this procedure:
CREATE PROCEDURE linkaccess AS SELECT OrderID FROM SERVER1.Northwind.dbo.Orderss
What do you think will happen if you try to create this procedure in SQL Server today? Depending on whether you have a linked server SERVER1 set up or not, you get one of these messages:
Msg 7202, Level 11, State 2, Procedure linkaccess, Line 2
Could not find server 'SERVER1' in sys.servers. Verify that the correct server name was specified. If necessary, execute the stored procedure sp_addlinkedserver to add the server to sys.servers.
Msg 7314, Level 16, State 1, Procedure linkaccess, Line 2
The OLE DB provider "SQLNCLI10" for linked server "SERVER1" does not contain the table ""Northwind"."dbo"."Orderss"". The table either does not exist or the current user does not have permissions on that table.
So when a stored procedure accesses a remote object, there is suddenly no longer any deferred name resolution! I find this ironic. Just like bulk-copy objects, this is a situation where I may prefer to not be alarmed about something missing, or at least not missing servers, at compile time. For instance, assume that as a DBA you have to apply a change script with a couple of stored procedures to your production database during a maintenance window. But you find that you cannot do that, because the stored procedures refer to a linked server which also is down for maintenance.
Thus, an idea here would be keep the current behaviour when strict checks are on, but change the behaviour when strict checks are off.
By the way, things are not any better with OPENQUERY:
CREATE PROCEDURE linkaccess2 AS SELECT * FROM OPENQUERY(SERVER1, 'SELECT OrderID FROM Northwind.dbo.Orderss')
OLE DB provider "SQLNCLI10" for linked server "SERVER1" returned message "Deferred prepare could not be completed.".
Msg 8180, Level 16, State 1, Line 1
Statement(s) could not be prepared.
Msg 208, Level 16, State 1, Line 1
Invalid object name 'Northwind.dbo.Orderss'.
There may be other types of objects or language constructs that I have overlooked and for which deferred name resolution could be an issue in one direction or another. I trust Microsoft to have good judgement to sort this out. Wait, what did I say? After the exposé above, how could I trust them with anything in this regard? :-) Anyway, if Microsoft wants my opinion for some feature I've overlooked, they are welcome to contact me.
Before I move on, I like to give credit to SQL Server Data Tools, SSDT. SSDT will alert you of many of the problems I have discussed in this section. SSDT understands to extract temp-table definitions and will give you an warning if you refer to a temp table that it does not know about, or if you misspell a column in a temp table. Unfortunately, though, it has no capability to deal with the situation where you create a temp table in one procedure to use it another.
We now leave the topic of deferred name resolution behind to turn to other areas where strict checks would make a difference. Consider this:
CREATE PROCEDURE get_order @OrderID varchar(5) AS SELECT OrderDate, CustomerID FROM Orders WHERE OrderID = @OrderID go EXEC get_order 110000
On SQL 7 and later the procedure is created, but the execution fails with:
Msg 245, Level 16, State 1, Procedure get_order, Line 2
Conversion failed when converting the varchar value '*' to data type int.
If this looks academic to you, I can tell you that this is from a real-world case where a colleague for some reason had declared a parameter as varchar(5) when it should have been int. It worked fine, until one day when the procedure was called with a six-digit number. That did not fit into varchar(5) and it went downhill from there.
Time for some nostalgia. Here is what SQL Server 6.5 has to say about the procedure above:
Server: Msg 260, Level 16, State 1, Procedure get_order, Line 2
Disallowed implicit conversion from datatype 'varchar' to datatype 'int' Table: 'tempdb.dbo.Orders', Column: 'OrderID' Use the CONVERT function to run this query.
And again, SQL 6.5 has it right. It's not that SQL 6.5 does not permit
implicit conversions at all, but it does not have implicit conversion between
numbers and strings. And I will have to confess that I just don't see the point
with this sort of implicit conversion. Sure, it permits me to say
PRINT '1'. But it also opens the door for unpleasant surprises.
Assume this table:
CREATE TABLE somedata(datakey varchar(10) NOT NULL PRIMARY KEY, whitenoise float NOT NULL DEFAULT rand(), filler char(4000) NOT NULL DEFAULT ' ') go INSERT somedata (datakey) VALUES ('123456') INSERT somedata (datakey) VALUES ('234567') INSERT somedata (datakey) VALUES ('9875222')
That is, in this table the key is a string, but the key values are mainly numeric. This query seems to run fine:
SELECT whitenoise FROM somedata WHERE datakey = 123456
Now add one more row and run the query again:
INSERT somedata (datakey) VALUES ('123456A') SELECT whitenoise FROM somedata WHERE datakey = 123456
The message is:
Msg 245, Level 16, State 1, Line 1
Conversion failed when converting the varchar value '123456A' to data type int.
The type conversion rules established in SQL 2000 say that when two types meet, the type with lowest precedence is converted to the other, if an implicit conversion exists. This has two nasty consequences in this case. As we have seen, as soon there is a non-numeric value in datakey, the query fails. But even if all data is numeric, all is not good. Since the varchar value is converted to int, any index on datakey is useless, and SQL Server must scan the table or the index. Try the query and look at the query plan.
By now, the reader understands what all this leads to: with strict checking on, there will be less implicit conversion permitted. It would be impractical to outlaw everything, since not all implicit conversions are harmful. For instance, converting from smallint to int is unproblematic in all contexts. I will have to admit that I have not considered each and every case, rather I will cover four generic cases which I cover below.
Before I go on, I like to make a distinction between the two situations where implicit conversions can occur: assignment and expressions. Assignment is all cases when data is written to something: a column, a variable or a procedure parameter. If the source is of a different data type than the target, the source is converted to the type of the target if there is an implicit conversion available. An expression is about anything else, for instance
col1 >= col2,
col2 + 91. As I mentioned above, SQL Server will in this situation convert the type with lower precedence to the other, again if an implicit conversion is available.
Note: I am under the impression that the relaxation of the type checks in SQL 7 were due to ANSI compliance. This mix of integers and strings is not unique SQL Server but is permitted with other products as well. But that does make it less harmful.
SQL Server 2017 has 33 basic types listed in sys.types. They can be divided into eight classes:
The basic idea is that when strict checking is in force, implicit conversion is not permitted from one class to another. However, the following implicit conversions would still be permitted:
That is, it is permitted to implicitly convert from String to several other data types. The rationale here is that you need to be able to work with literals, and for instance there is no date literal in T‑SQL. Well, SQL 2012 added datefromparts(), datetimefromparts() etc, so you could argue that there is no longer any need for implicit conversion from String to Date/time. But these functions are a little verbose. And, maybe more importantly, there is probably a lot of code out there that relies on this implicit conversion. (Yes, there is also a lot of code that relies on implicit conversion from Strings to Numeric. I see that a lot on the Transact-SQL forums. People mix data types and then they get problems at run-time they don't understand, because SQL Server did not stop them earlier.)
It is not equally compelling to have implicit conversion from Date/time to String or Uniqueidentifier to string, but neither is there any major harm, which is why I have put these parentheses. (XML, sql_variant and CLR types to String are not listed since there are no such implicit conversions today.)
It is also permitted to implicitly convert to sql_variant; that's kind of the essence of that data type. (XML and CLR types are not included, since they cannot be stored in sql_variant.)
This means that quite a few implicit conversions permitted today would be disallowed
with strict checks on; even
some that were permitted in SQL 6.5, for instance binary to numeric. The most
controversial ban may be integer to date/time (this was not permitted in
6.5 either). And, in fact Microsoft took that step in SQL 2008 with the new date/time data types. You can say
SELECT @date + 1, but only if @date is datetime or smalldatetime, but not if it is datetime2 or date. I know some people think this is useful, but I only find it corny. And what do you think about
SELECT @date = 2008-05-30? That is
not an uncommon mistake.
When there is a mix of types from different classes, there is no difference between assignment and expressions. The same rules for which implicit conversions that are permitted apply for both.
When I originally wrote this article, I said:
This is where it gets more difficult, and I will be very loose here. One possibility would be that any conversion that could incur loss of information would require explicit conversion with strict checks: from nvarchar to varchar, from float to int, from varchar(23) to varchar(8). But that could be a bit too strict even for me in some situations. So I pass for now.
However, some time back I was approached by Nicholas Chammas, and he convinced me that there is no reason why this should not be covered by strict checks. That is, if an implicit conversion could lead to loss of information, this should yield an error when strict checks are in effect. In practice, this only concerns assignment, since in an expression the shorter type is always converted to the longer type.
Let's look at some examples.
CREATE TABLE abc(a varchar(5) NOT NULL) go CREATE PROCEDURE insert_value @a varchar(10) AS INSERT abc(a) VALUES (@a) go EXEC insert_value 'Too long!'
Today, without strict checks, this will yield the run-time error:
Msg 8152, Level 16, State 14, Procedure insert_value, Line 2
String or binary data would be truncated.
The statement has been terminated.
Surely, it is much better to be told at compile-time that the procedure is unsafe?
SQL Server is not consistent. If you say:
DECLARE @a varchar(5), @b varchar(10) SELECT @b = 'Too long!' SELECT @a = @b
There is no error, but @a will be assigned the value Too l. But under strict checks this implicit conversion would not be permitted. If you wanted to accept the data loss you would need to use cast, convert or substring to state your intention explicitly. (If you wonder why SQL Server is not consistent, the background is that up SQL Server 6.0, truncation was always silent. To be ANSI-compliant, Microsoft added an option for raising an error when storing values in a column, and this option is today the norm. I guess they did not make the same change for variables for compatibility reasons.)
The same applies if you try to assign nvarchar to varchar:
DECLARE @v varchar(20), @n nvarchar(20) SELECT @n = N'Lech Wałęsa' SELECT @v = @n SELECT @v
Unless you have a collation based on a code page that supports Polish, the output is
Note that the two middle characters in the last name have been altered. Since this is loss of information, this should yield an error when strict checks are on. You would need to explicitly use cast or convert.
What about numbers? The same principles should apply, although there is some fine print to sort out. Here are some clear-cut cases, which all should yield an error with strict checks on:
DECLARE @i int, @s smallint, @f float, @r real SELECT @s = @i, @i = @f, @r = @f
You would need to use cast or convert to avoid the error. These assignments should be legal:
SELECT @i = @s, @f = @i, @f = @r
With the decimal data types, it's a little messy Strictly(!) speaking both assignments below should be illegal:
DECLARE @small_dec decimal(5, 0), @large_dec decimal(18, 0), @i int SET @small_dec = @i SET @i = @large_dec
I leave it to Microsoft to make the ultimate decision here, as I do with the various date/time data types which also invites to similar considerations.
There is one situation where the conversion due to data-type precedence can cause performance problems: if the expression is a comparison operator that appears in a WHERE or ON clause, and a column is implicitly converted in a way that precludes efficient use of any index on the column. With strict checks in effect, such implicit conversions would not be permitted.
To clarify. This should always be permitted:
SELECT @nvarchar = 'somevarcharstring' UPDATE tbl SET nvarcharcol = varcharcol SELECT col FROM tbl WHERE nvarcharcol = @varcharval
In these three examples a varchar value is implicitly converted to nvarchar. This should always be permitted, as there is no loss of information and there is no performance issue. But this is bad:
SELECT col FROM tbl WHERE varcharcol = @nvarcharval SELECT col FROM tbl WHERE varcharcol = N'nvarchar' SELECT a.col FROM tbl1 a, tbl2 b WHERE a.varcharcol = b.nvarcharcol
In all these queries, the varchar column gets converted to nvarchar. If the column has a Windows collation, the index can be still seeked, but in a less efficient way. With an SQL collation, the index is useless. The rule should not be dependent on the collation, but the worst case should be considered. Thus, all these queries would reward an error message when strict checks are in effect. Note that there should be an error, regardless whether the column is indexed or not. (As an index may be added in the future.)
Here is another good query:
SELECT col FROM tbl WHERE tinyintcol = @intvar SELECT col FROM tbl WHERE intcol = 11000.0
According to the conversion rules, tinyint should be converted to int, but this is not how it looks in the query plan. The same is true for the second query.
This query is less clear-cut:
SELECT a.col FROM tbl1 a, tbl2 b WHERE a.tinyintcol = b.floatcol
When I fooled around with a query like this, I got an implicit conversion on tinyintcol, if tbl1 was the table that was scanned, and thus the implicit conversion was harmless. But the rules should of course be independent of the query plan. This case needs further investigation.
It may be worth pointing out that the error message in this case should not say Implicit conversion ... is not allowed. Use the CONVERT function to run this query, but rather encourage the programmer to avoid the type clash altogether.
Consider this batch:
EXEC sp_addtype thistype, 'varchar(10)' EXEC sp_addtype thattype, 'varchar(10)' go CREATE TABLE domaintest (a thistype NOT NULL, b thattype NOT NULL) go SELECT * FROM domaintest WHERE a = b
From all versions from SQL 7 and on, this passes. But SQL 6.5 objects:
Server: Msg 305, Level 16, State 1, Line 1
The column 'a' (user type:thistype) is joined with 'b' (user type:thattype). The user types are not compatible: user types must be identical in order to join.
If you have used a very strongly typed language like Ada, this is perfectly logical. Now, SQL 6.5 was quite inconsistent. This restriction applied to joins only. It was OK to mix user-defined types in assignments, when comparing variables or even in foreign-key constraints. So I can understand why Microsoft dropped this rule in SQL 7.
I am not going to call for a change with regards to user-defined types. This feature is probably not used widely enough to warrant that. But hopefully one day (already in the next version of SQL Server???), Microsoft will add real domains from ANSI SQL. I don't know what conversion rules that ANSI mandates, but my suggestion is that with strict checks on, there would be no implicit conversion whatsoever between values from different domains. A customer id and an order id may both be integer, but if you are joining them you are doing something wrong.
DECLARE @str varchar, @dec decimal SELECT @str = 'My string', @dec = 12.2234
What value does @str and @dec have now? That's right, M and 12. But it is not uncommon to see questions on the SQL Server forums from people who had different expectations. And one can hardly blame them. A default of 1 for a variable-length string is just plain silly. And while maybe not silly, the defaults of 18 and 0 for decimal are not obvious. And more importantly, these people may be accustomed from other environments where you don't specify precision and scale for decimal at all, for instance .Net.
So with strict checks in force, there would be no default length for char, nchar, varchar, nvarchar, binary and varbinary, but you must always specify it explicitly. Nor would there be any default precision or scale for decimal and numeric. And for consistency, scale should also be mandatory for time, datetime2 and datetimeoffset.
However we need one exception. Consider:
SELECT cast(intcol AS varchar)
In this marvellously consistent language known as Transact-SQL, the default here is varchar(30), not varchar(1). So in this specific example, there is no risk for truncation. And it is likely that there is a lot of code out here which casts numbers or datetime values to string in this way. There is no reason to raise an unnecessary hurdle for the adoption of strict checks. So it should be alright to leave out the length on cast and convert – as long as there is no risk for truncation. Here are three examples that all would yield an error:
SELECT cast(sysdatetimeoffset() AS varchar) SELECT cast(newid() AS char) DECLARE @v varchar(50); SELECT cast(@v AS nvarchar)
If your stored procedure calls other procedures, the sole check at compile time is that the procedures exist, and as discussed above, you only get a warning if a procedure is missing, not an error. If you have left out any mandatory parameters, or specified a non-existing parameter, you will not be told until run-time. The same goes if you specify OUTPUT for a parameter that is not an output parameter. And if you leave out OUTPUT in the EXEC command for an output parameter, you don't even get an error at run-time!
In contrast, if your stored procedure calls a user-defined function, you get errors for missing or superfluous parameters already at compile-time. With one exception: if you call a scalar UDF through EXEC, the behaviour is the same as when you call a stored procedure.
Obviously there is room for improvements. With strict checks in force the following would apply to the EXEC command when you submit an SQL batch:
In all these cases, when I say "error", I do of course mean "compile-time error".
These checks can obviously not be performed when the procedure name is
specified through a variable, for instance
EXEC @spname. Special
procedures like sp_executesql will require some extra consideration.
Should these checks be performed for calls to procedures in linked servers? I think so. At least, it should be consistent with how references to tables in linked servers are handled. And, as we have seen, they are checked even today.
I should add that SSDT performs some of these checks, for instance it warns you for superfluous parameters.
CREATE PROCEDURE print_this @this varchar(20) AS PRINT @this go DECLARE @that varchar(20) SELECT @that = 'what?' EXEC print_this @this = that
What would you expect this to result in? The actual output is
The reason for this is that in EXEC statements there is no requirement to put a string literal in quotes, as long as the literal conforms to the rules for regular identifiers. This is the rule that permits us to write:
without putting the database name in quotes. And while this may be handy in an ad-hoc session, it is only a source for error in a programming environment. In the example above, the intention was presumably to pass the variable @that to the stored procedure.
Thus, with strict checks in force, it would be an error to have a string literal without quotes in an EXEC statement inside a stored procedure. (It could still be permitted in the top-level scope, to facilitate the use of system procedures.)
If you say
SELECT TOP 20 col1, col2 FROM tbl
SQL Server is free to return any twenty rows in the table. Sometimes this is what you want – you only want 20 rows and you don't care which rows. But it could also be that you inadvertently left out the ORDER BY clause. Or you are in the false belief that it is not needed but you will always get the 20 "first" rows, whatever your idea of "first" may be.
Therefore, it seems that it would be a good idea to make ORDER BY compulsory with TOP in strict-checks mode. If you really don't care about the order, you need to specify this explicitly:
SELECT TOP 20 col1, col2 FROM tbl ORDER BY (SELECT NULL)
The same rule applies already today to the row_number() function: you must have an ORDER BY, but you can use a constant subquery if you don't care about the order.
Consider this INSERT statement:
INSERT archived_orders(order_id, order_date, invoice_date, customer_id, employee_id) SELECT order_id, invoice_date, order_date, customer_id, employee_id FROM orders WHERE order_date < @12monthsago
As you see, most likely there is an error in this INSERT statement: the two date columns have been swapped. This may seem like an error that is simple to catch with a visual inspection. However, imagine that the INSERT statement involves 50 columns and the swapped columns are in the middle.
Here is another example:
INSERT sometbl(Albert, Boris, Cesar, David, Eric, Fiona, Greta, Heinrich) SELECT Albert, Boris, Cesar, David Eric, Fiona, Greta, Heinrich, extra FROM othertable WHERE ...
At first glance, you may think this that this will not compile, but fail due to a mismatch in the number of columns. However, there are two errors in the SELECT statement. Not only is there an extraneous column at the end, but there is also a comma missing after
David. You may think that
David Eric, i.e. two subsequent identifiers, is a syntax error, but it isn't. This is equivalent to
David AS Eric. Thus, there is some chance that the INSERT statement will run successfully, and cause incorrect data to be inserted. Again, imagine an INSERT statement with 50 columns where you may not notice the extraneous column, even less the missing comma.
It seems to me that the INSERT statement was not really designed for industrial-strength programming, but it is the way it is and it will not change. But maybe we could find something within the realm of strict checks to increase our confidence in our long INSERT statements?
A very strict rule, but also largely unrealistic is that column names in the INSERT and SELECT lists must match exactly. That is, the following would be illegal:
INSERT tbl (a, b, c, d) SELECT a, x, 1, coalesce(d, 0) FROM src
You would have to write it as:
INSERT tbl (a, b, c, d) SELECT a, x AS b, 1 AS c, , coalesce(d, 0) AS d FROM src
I would suspect that a minority of the INSERT-SELECT statements out there complies to this rule. A more realistic rule might be this: if any source column has an alias, all names in the SELECT list must match the INSERT list. If a column is aliased, it is the alias that must match. That is, the two INSERT statements above would both be legal, but this would be illegal:
INSERT tbl (a, b, c, d) SELECT a, b AS myownalias, 1 AS c, coalesce(d, 0) FROM src
INSERT tbl (a, b, c) SELECT wrongcolumn, x AS b, 1 AS c FROM src
One could argue that only explicitly aliased columns should be checked. But once this rule goes into effect and gains popularity, it would surely be considered best practice to alias all source columns with a different name from the target column, so inadvertently leaving out an alias is something you want to be alerted of. And moreover, once you see that one column has an alias, you can inspect the SELECT query, and do not really have to consider the INSERT list and look for matching errors.
What if a column in the SELECT list is a constant or an expression? Does it need to have an alias? That is, is this legal or not:
INSERT tbl (a, b, c, d) SELECT a, x AS b, 1, coalesce(d, 0) FROM src
I have not acquired a firm opinion, but I lean towards that it should be required for expressions but maybe not for constants (since
NULL AS col could be perceived as bulky). Thus, the above would be illegal, but the below would pass:
INSERT tbl (a, b, c, d) SELECT a, x AS b, 1, coalesce(d, 0) AS d FROM src
Here I have only covered the INSERT statement, but these pairs of lists that you need to keep aligned appear in several places in SQL:
Definitions of CTEs and views, e.g.
CREATE VIEW v (a, b, c) AS ... Although you could argue in this case the column list is optional, so if the programmer leaves it out there is no risk for error. Then again, the programmer may like to add it for clarity. Better then to check it, so that the list does not introduce an error.
Row constructors. This is a feature in ANSI SQL that T‑SQL does not support, but which is high on many wish-lists. With row constructors you would be able to write:
UPDATE tbl SET (col1, col2, col3) = (SELECT col1, col2, col3 FROM ...)
If row constructors are added to SQL Server, the same checks should apply as to the INSERT statement, including variable assignment:
SET (@a, @b, @c) = (SELECT alfa, beta, cesar AS c FROM tbl)
Since one column is aliased, all columns must match the variable names, save the leading @.
The set operators UNION, INTERSECT and EXCEPT. With these, the first SELECT determines the column names of the statement and any aliases in the subsequent SELECT are ignored. The SELECT lists for such queries certainly have potential for alignment errors, not the least if the queries have complex FROM and WHERE clauses, so that the SELECT lists are far apart. Thus, this would be legal with strict checks:
SELECT a, b, c FROM tbl1 UNION ALL SELECT e, f, g FROM tbl2 SELECT a, b, c FROM tbl1 UNION ALL SELECT e AS a, f AS b, g AS c FROM tbl2
But this would be illegal:
SELECT a, b, c FROM tbl1 UNION ALL SELECT e AS a, f, g FROM tbl2
Here is one that SQL Server MVP Razvan Socol ran into. He happened to write:
DECLARE @Something datetime ... UPDATE SingleRowTable SET @Something=NULL
Which is perfectly legal, but of course wasn't what he intended. The @ was a slip on his part. And something that SQL Server could have alerted him about. Generally, while UPDATE permits you to assign variables, it could only make sense if you assign it an expression that includes at least one column in the target table for the UPDATE. (Else you could just well use a SELECT.) And even then it's a bit fishy; there are a few cases when you roll your own id columns where it comes in handy, but they are not that common. So I could even go as far as arguing that variable assignment in UPDATE should not be permitted at all in strict mode. The few cases where it's useful have to be weighed against the many more cases it's a programming error.
There are people who argue that all column references in a query should be prefixed, period. I can sympathise with the idea, but I will have to admit that I much prefer the version to the left in the queries below:
SELECT OrderID, CustomerID, OrderDate SELECT O.OrderID, O.CustomerID, O.OrderDate FROM Orders FROM Orders O WHERE EmployeeID = 19 WHERE O.EmployeeID = 19
But there are certainly cases where not using column prefixes can lead to unpleaseant surprises, and I will discuss one special case and then move on to the general case.
CREATE TABLE Orders(ordOrderID int NOT NULL PRIMARY KEY, ordCustomerID int NOT NULL, ...) CREATE TABLE OrderDetails (detOrderID int NOT NULL, detRowNo smallint NOT NULL, ...) SELECT ... FROM Orders WHERE ordOrderID IN (SELECT ordOrderID FROM OrderDetails WHERE detProductID = 76)
The error here is that the subquery has a column from the outer table in the SELECT list. Perfectly legal, but not that meaningful. More than one have been bitten by this error and posted to SQL forums, thinking that they have hit a bug in SQL Server. They haven't, but with strict checks we could help them to detect their mistake earlier. I have two suggestions:
Both protects against the mishap above in the SELECT list on their own, but you can be coding half asleep and use the wrong alias, in which case the second rule saves you. The first rule, on the other hand, picks up possibly unintended use of columns from the outer tables elsewhere in the subquery.
CREATE TABLE #temp (Product_id int NOT NULL PRIMARY KEY, Turnover float NOT NULL) ... SELECT Products.Id, Product_name, Category_name, Whizbang, Turnover FROM Products JOIN Categories ON Category_id = Categories.Id JOIN #temp ON Products.Id = Product_id
Here the programmer has computed the turnover for a product in a temp table, and returns the data together with som data about the product and the product category in a query. Beside the two Id columns, all column names are unique, and the programmer being of a lazy nature, thinks that he does not need to use more column prefixes than absolutely required.
We can start with the observation that queries like this one are difficult to read for the outsider who don't know the tables. Since the temp table is declared in the same procedure, we can tell where Turnover comes from. But Whizbang? Is that in Products or Categories? While irritating, this alone is not sufficient to warrant any compile-time checks in my opinion. There is however a second problem: this code could fail at some point in the future. One day the DBA decides to add a Turnover column to the Products table. The code above will now fail to compile with
Msg 209, Level 16, State 1, Line 1
Ambiguous column name 'Turnover'.
The DBA might be prepared that code could break if he drops a column, but if he adds a column he is likely to have his guard down. The above example is apparently from an outright sloppy and indifferent programmer, but even a good programmer who knows to prefix his column may forget it from time to time.
Therefore it would be a good idea if strict checks would trap column references that could become ambiguous in the future. My suggestion for a firm rule is this one: if more than one table source is visible in a certain place in a query, all columns must be prefixed with an alias or the table name. This rule also covers the situation in the previous section, where there is no risk for ambiguity but well for confusion. To illustrate the rule, here is an example queries where column references what would be errors under strict checks are marked in red:
; WITH CTE AS ( SELECT a, b, c, rowno = row_number() OVER(ORDER BY a) FROM tbl ) SELECT e, f, a, cnt FROM CTE JOIN tbl2 ON a = k JOIN (SELECT d, COUNT(*) AS cnt FROM someothertbl GROUP BY d) AS aggr ON d = p WHERE rowno = 1 AND NOT EXISTS (SELECT * FROM tbl3 WHERE g <> b)
The CTE and the derived table are OK, because there is only table source visible in these. (Inside a derived table, the tables in the outer query are not visible.)
We now enter the area of somewhat more speculative suggestions. We start with looking at cursors and compile-time checks for something it is by its definition dynamic. To wit, despite that the statement reads DECLARE CURSOR, it's an executable statement, and as a consequence of this, there is no compile-time check whatsoever of cursors. If you misspell the cursor name in the OPEN, FETCH, CLOSE or DEALLOCATE statements it will not bite you until run-time. The same is true if your FETCH statement does not match the cursor declaration.
And of course, in the general case, such checks are impossible. It's
perfectly legal to declare a cursor in one scope and access it from another. But
how common is that? In my experience, a cursor is almost always created and used
in one single stored procedure. The most common exception is probably when you set up the cursor
in dynamic SQL, because your conditions for the cursor are dynamic. Thus, it
seems to be a good trade-off to add checks for cursors in strict mode, and pay
the price of having to add
/* NOSTRICT */ in the odd case.
With strict checks in force, SQL Server would extract cursor declarations and compare these with other cursor statements:
If you use cursor variables, you get the second rule for free, but not the others. The remaining checks could be implemented for local cursor variables. For cursor parameters (yes such exist!), there would be no checks.
One more small thing with cursors, although unrelated to the above: it could be worth considering whether it should be permitted to leave the cursor type unspecified in strict mode. That is, in strict mode, you must specify the cursor as INSENSITIVE, STATIC, KEYSET, FAST_FORWARD or DYNAMIC. The default cursor type today is apparently DYNAMIC, which is a very poor choice. Most of the time, people probably think in the mind-set of a static cursor. For the same reason, LOCAL would be required. (And GLOBAL an error?)
With the old ANSI syntax for the FROM clause, it was a bit too easy to write things like:
SELECT a.col1, b.col2 FROM a, b, extra WHERE a.keycol = b.keycol AND b.somecol = @value
That is, you could accidently have an extra table in your FROM clause that formed a Cartesian product with the rest of the query. With the newer syntax with JOIN and ON, the risk for this debacle is removed since you have to explicitly specify CROSS JOIN if you want it. Unfortunately, you can still do this mistake:
SELECT a.col1, b.col2 FROM a JOIN b ON a.keycol = a.keycol
I don't know about you, but I do it often enough to want SQL Server to tell me about it.
The basic idea is that in strict mode, the ON clause must refer to the table source that the ON clause is attached to, and at least one of the preceding table sources. But we need to consider two complications: nested joins and multi-column joins before we can make a firm rule out of this.
It's legal to nest joins. Most often this is done with outer joins. Here is an example:
SELECT ... FROM a LEFT JOIN (b JOIN c ON b.col1 = c.col1) ON a.col2 = b.col2
How would this work with strict checks? Let's look at a list of cases:
LEFT JOIN (b JOIN c ON b.col1 = c.col1) ON a.col2 = b.col2 LEFT JOIN (b JOIN c ON b.col1 = c.col1) ON a.col2 = c.col2 LEFT JOIN (b JOIN c ON b.col1 = c.col1) ON b.col2 = c.col2 LEFT JOIN (b JOIN c ON b.col1 = b.col1) ON a.col2 = b.col2 LEFT JOIN (b JOIN c ON a.col1 = c.col1) ON a.col2 = b.col2
The first two both makes perfect sense, and should not result in any error. The third on the other hand looks spooky. Somehow the table a is left free-floating. The fourth is a breach against the basic idea, as only one table is included in the condition. The fifth is an error in SQL Server today: in a nested JOIN condition, table sources outside the nested join are not visible; similar to a derived table.
Thus, for the innermost join in a set of nested joins we don't need any modification of the basic rule. But for the joins further out, "the table source the ON clause is attached to" refers to the nested join, that is b JOIN c in the example above.
The next problem to consider is multi-column conditions. This a likely to be a goof:
SELECT a.col1, b.col2 FROM a JOIN b ON a.keycol1 = b.keycol1 AND b.keycol2 = b.keycol2
It passes the basic rule, so we need to extend it a bit. ON clauses typically consist of a number of AND factors. So the rule could be extended to all AND factors? Not so fast! This is perfectly OK:
SELECT a.col1, b.col2 FROM a LEFT JOIN b ON a.keycol = b.keycol AND b.startdate = @startdate
So the rule needs to be modified to: each AND factor must include a column from the table source the ON clause is attached to, and either a preceding table source or a variable/constant. And at least one AND factor must refer to a preceding table source.
This still renders some legit queries as incorrect in strict mode, for instance:
SELECT a.col1, b.col2 FROM a LEFT JOIN b ON a.keycol = b.keycol AND b.startdate = b.enddate
I have never seen this, but who knows, maybe someone would like to do:
SELECT a.col1, b.col2 FROM a LEFT JOIN b ON a.keycol = b.keycol AND @include_b = 1
I contemplated these two cases for a while, and considered variations to avoid that they
would be flagged as errors in strict mode, but I arrived at the conclusion that
these sorts of conditions are rare enough. If the programmer wants to do this, he
needs to add
/* NOSTRICT */ to suppress the error message. I think this is OK,
as long the checks more often help the programmer from doing silly goofs than
they get in the way for him.
Here all conditions were plain = and I only talked about AND. There is not really any difference to other operators. This condition is alright with strict checks:
SELECT a.col1 FROM a JOIN b ON datediff(DAY, a.somedate, b.somedate) <= 7
What this means can be disputed, but it fulfils the rule we've set up here, and I see no reason to change it. The same apply to more complex conditions that include CASE expressions. As for subqueries, if we have:
SELECT a.col1, b.col2 FROM a JOIN b ON b.col = (SELECT ...)
The subquery must refer to a column from a to be accepted in strict mode.
The reader may think that OR terms should be handled similarly to AND factors, but OR is so much less common, that I don't think it's worth making any
it more complex by adding rules for OR. Rather OR is best considered to be an operator just like
=. So this will pass, even if it's likely
to be a goof:
SELECT a.col1, b.col2 FROM a JOIN b ON a.keycol1 = a.keycol1 OR b.altkey = b.altkey
There are columns from both tables in the single AND factor, so this passes.
I don't suggest any particular checks for WHERE clauses. Well, maybe really
obvious things like
WHERE a.col = a.col should raise an error, but
as soon as you involve expressions, it gets more difficult to tell the goofs
from the real intentions. For instance, this makes perfect sense on a case-insensitive
UPDATE tbl SET col = upper(col) WHERE col != upper(col)
Going back to the first example:
SELECT a.col1, b.col2 FROM a, b, extra WHERE a.keycol = b.keycol AND b.somecol = @valueOf course, it would not be a bad idea to flag this as an error. More precisely, if a table appears in the FROM clause after a comma, but never appears in the WHERE clause this is an error. But I never use this join style anymore, so for me this is not that compelling. :-)
Here is yet a suggestion of a more speculative nature and which requires further investigation. But I feel that here is a great potential to improving the quality of SQL Server code world-wide by alerting programmers of errors that are due to sloppiness or lack of understanding of key features.
The cardinality errors I have in mind are contexts when at most one row should be returned, but where there is no compile-time guarantee that this is the case. In this section, most of the examples are based on these tables:
CREATE TABLE header (id int NOT NULL, a int NOT NULL, b int NOT NULL, CONSTRAINT pk_header PRIMARY KEY (id) ) CREATE TABLE lines (id int NOT NULL, rowno int NOT NULL, b int NOT NULL, CONSTRAINT pk_lines PRIMARY KEY(id, rowno), CONSTRAINT fk_lines_header FOREIGN KEY (id) REFERENCES header(id) ) go INSERT header (id, a) VALUES (1, 112, 90000) INSERT lines (id, rowno, b) VALUES (1, 1, 12), (1, 2, 14), (1, 3, 16) go
All the following statements have a cardinality error.
SELECT @b = lines.b /*1*/ FROM lines JOIN header ON header.id = lines.id WHERE header.id = 1 SET @b = (SELECT lines.b /*2*/ FROM lines JOIN header ON header.id = lines.id WHERE header.id = 1) SELECT header.id, header.a, (SELECT lines.b FROM lines WHERE lines.id = header.id) /*2*/ FROM header SELECT header.id, header.a, header.b /*2*/ FROM header WHERE header.b = (SELECT lines.b FROM lines WHERE lines.id = header.id) UPDATE header /*1*/ SET b = lines.b FROM header JOIN lines ON header.id = lines.id UPDATE header /*2*/ SET b = (SELECT lines.b FROM lines WHERE header.id = lines.id) WHERE EXISTS (SELECT * FROM lines WHERE header.id = lines.id) MERGE header /*3*/ USING lines ON header.id = lines.id WHEN MATCHED THEN UPDATE SET b = lines.b ;
Today, the behaviour in SQL Server is that all pass compilation. At run-time, the statements marked 1 completes successfully, however the result is non-deterministic. That is, @b and header.b could be set to any of 12, 14 and 16. The statements marked 2 all result in this error:
Msg 512, Level 16, State 1, Line 1
Subquery returned more than 1 value. This is not permitted when the subquery follows =, !=, <, <= , >, >= or when the subquery is used as an expression.
The statement has been terminated.
Finally, the MERGE statement has its own error message:
Msg 8672, Level 16, State 1, Line 1
The MERGE statement attempted to UPDATE or DELETE the same row more than once. This happens when a target row matches more than one source row. A MERGE statement cannot UPDATE/DELETE the same row of the target table multiple times. Refine the ON clause to ensure a target row matches at most one source row, or use the GROUP BY clause to group the source rows.
But why wait until run-time? Given the table definitions, we could see with our own eyes that the statements are problematic. On the other hand, we can easily tell that these are safe:
SELECT @b = b FROM header WHERE id = 1 SET @b = (SELECT b FROM header WHERE id = 1) SELECT lines.id, lines.rowno, lines.b, (SELECT header.a FROM header WHERE lines.id = header.id) FROM lines UPDATE lines SET b = header.b FROM header JOIN lines ON header.id = lines.id UPDATE lines SET b = (SELECT header.b FROM header WHERE header.id = lines.id) WHERE EXISTS (SELECT * FROM header WHERE header.id = lines.id) MERGE lines USING header ON header.id = lines.id WHEN MATCHED THEN UPDATE SET b = header.b ;
Since there is a primary key on id, the join or subquery on header can return at most one row. So that is the basic rule: when there is a condition which does not include the complete primary key (or any other unique index), SQL Server should raise an error when strict checks are active for:
Some qualifications are called for. In these queries the primary key is implicit in the CTE:
WITH CTE AS ( SELECT id, MIN(b) AS b FROM lines GROUP BY id ) UPDATE header SET b = CTE.b FROM header JOIN CTE ON header.id = CTE.id
WITH CTE AS ( SELECT id, b, row_number() OVER(PARTITION BY id ORDER BY rowno) AS rn FROM lines ) UPDATE header SET b = CTE.b FROM header JOIN CTE ON header.id = CTE.id WHERE CTE.rn = 1
There is one situation where there is no need for any key to be present, and that is if you use TOP 1. This should be legal:
SELECT TOP 1 @b = lines.b FROM lines JOIN header ON header.id = lines.id WHERE header.id = 1 ORDER BY (SELECT NULL)
As discussed above, if you don't really care what value @b is assigned to, you need to state it explicitly.
Unique filtered indexes should also be considered. Here is an example. For all orders that have not been assigned an explicit address, we need to update the orders with a default address:
UPDATE orders SET address1 = adr.address1 address2 = adr.address2, ... FROM orders ord JOIN customeraddresses cad ON ord.customerid = cad.customerid AND cad.isdefaultaddress = 1 JOIN addresses adr ON cad.adrid = adr.adrid WHERE ord.address1 IS NULL
The primary key on customeraddresses is (customerid, adrid), so the statement is potentially non-deterministic. However, this query should pass under strict checks if and only if there is a unique filtered index on
customeraddress (customerid) WHERE isdefaultaddress = 1
since in this case we can deduce that the UPDATE is unambiguous. This could be further extended to indexed views and indexed computed columns, but I leave it to Microsoft to explore that ground.
One more thing needs to be said about UPDATE FROM. Consider:
UPDATE header SET b = 0 FROM header JOIN lines ON header.id = lines.id
While this UPDATE statement (logically) hits the same target row many times, this is still deterministic: header.b will be set to 0 for the header rows where there are line rows. So there should not be any compile-time error here, strict checks or not. In the same vein, this is also safe:
MERGE header USING lines ON header.id = lines.id WHEN MATCHED THEN UPDATE SET b = 0 ;
However, today this yields the run-time error above, so for this reason, SQL Server should raise a compile-time error with strict checks to help the programmer to be on the front line. (Or the run-time error should not appear in this situation.)
Before I close this section, I like to make one more comment on exactly how useful these checks could be. A very common error that comes up again and again on the forums is:
CREATE TRIGGER bad_trigger ON tbl FOR INSERT AS DECLARE @col SELECT @col = col FROM inserted ...
For whatever reason, inexperienced SQL Server programmers think that their trigger only need to consider single rows. Maybe because they have not heard of multi-row operations, maybe they come from Oracle where per-row triggers is the norm. Whatever, strict checks would tell them up front that they doing something wrong.
By adding a new feature, in this text called "strict checks"; Microsoft can help programmers to find silly and stupid errors early, and thereby help them to be more productive and produce a work of higher quality. In this text, I've discussed a number of things that could be covered by strict checks. I noted initially, that adding strict checks for some things in one release, and adding further checks in a later release will cause compatibility problems. Therefore, there is reason for a big bang when it comes to the box product. (But as noted the possibility to have things in preview in Azure opens for the possibility to expose checks gradually as they are implemented.)
Nevertheless, some priorities may be in order. By far the most important is the death of deferred name resolution. I have ordered the checks roughly in priority order, but I have also considered the expected difficulty to implement the check. The last item, fixing cardinality errors, looks very interesting to me, but it is also one that requires careful study to get right.
I suspect that Microsoft feels that I have given them more than a mouthful with this list. Nevertheless, if you have further suggestions, please feel free to drop me a line at email@example.com. If I agree with you, I may add the suggestion to the article. If not, you can always file your own suggestion on the SQL Server feedback site, https://feedback.azure.com/forums/908035-sql-server.