Error Handling in SQL 2000 – a Background

An SQL text by Erland Sommarskog, SQL Server MVP. Last revision 2009-11-29.


This is one of two articles about error handling in SQL Server 2000. This article focuses on how SQL Server – and to some extent ADO – behave when an error occurs. The other article, Implementing Error Handling with Stored Procedures, gives advice for how you should check for errors when you write stored procedures. Logically, this article is part one, and Implementing... is part two. However, you can read the articles in any order, and if you are relatively new to SQL Server, I recommend that you start with Implementing.... The article here gives a deeper background and may answer more advanced users' questions about error handling in SQL Server.

Note: this article was written for SQL 2000 and earlier versions. All I have for SQL 2005 is unfinished article with a section Jumpstart Error Handling. The content in this article is to some extent applicable to SQL 2005 as well, but you will have to use your imagination to map what I say to SQL 2005. The article includes a short section on TRY-CATCH. I hope to produce a complete article for error handling in SQL 2005 later on.

Table of Contents:

   Introduction
   The Basics
      The Anatomy of an Error Message
      How to Detect an Error in T-SQL – @@error
      Return Values from Stored Procedures
      @@rowcount
      @@trancount
      More on Severity Levels
   What Happens when an Error Occurs?
      The Possible Actions
      When Does SQL Server Take which Action?
         Connection-termination
         Scope-abortion
         Statement-termination and Batch-abortion
      Trigger Context
      Errors in User-Defined Functions
   Control Over Error Handling
      SET XACT_ABORT
      ARITHABORT, ARITHIGNORE and ANSI_WARNINGS
      RAISERROR WITH NOWAIT
      Duplicates
      Using Linked Servers
      Retrieving the Text of an Error Message
      TRY-CATCH in SQL 2005
   Client-side Error Handling
      DB-Library
      ODBC
      ADO
      ADO .Net
   Acknowledgements and Feedback
   Revision History

Introduction

In many aspects SQL Server is a very good DBMS that permits you implement powerful solutions with good performance. However, when it comes to error handling... To be blunt: error handling in SQL Server is poor. It is a patchwork of not-always-so-consistent behaviour. It's also weak in that you have fairly little control over error handling, and for advanced error handling like suppressing errors or logging errors, you must take help from the client-side. Unfortunately, depending on which client library you use, you may find that the client library has its own quirks, sometimes painting you into a corner where there is no real good solution.

In this article, I will first look at what parts an error message consists of, and how you can detect that an error has occurred in T-SQL code. Next, I describe the possible actions can SQL Server can take in case of an error. I then proceed to describe the few possibilities you have to control SQL Server's error handling. Finally, there is a section on how the different client libraries from Microsoft behave, with most of the focus on ADO and ADO .Net.

General disclaimer: whereas some information in this text is drawn from Books Online and other documentation from Microsoft, a lot of what I say is based on observations that I have made from working with SQL Server, and far from all of this is documented in Books Online. Therefore, you should be wary to rely on a specific behaviour like "this error have this-and-this effect", as it could be different in another version of SQL Server, even different between service packs.

The Basics

The Anatomy of an Error Message

Here is a typical error message you can get from SQL Server when working from Query Analyzer.

Server: Msg 547, Level 16, State 1, Procedure error_demo_sp, Line 2
UPDATE statement conflicted with COLUMN FOREIGN KEY constraint 'fk7_acc_cur'.
The conflict occurred in database 'bos_sommar', table 'currencies', column 'curcode'.
The statement has been terminated.
Note: Under Tools->Options->Connections, I have checked Parse ODBC Message Prefixes.

The error information that SQL Server passes to the client consists of several components, and the client is responsible for the final interpretation of the message. These are the components that SQL Server passes to the client.

Message number – each error message has a number. You can find most of the message numbers in the table sysmessages in the master database. (There some special numbers like 0 and 50000 that do not appear there.) In this example, the message number is 547. Since most interesting messages are errors, I will also use the term error number. Message numbers from 50001 and up are user-defined. Lower numbers are system defined.

Severity level a number from 0 to 25. The stort story is that if the severity level is in the range 0-10, the message is informational or a warning, and not an error. Errors resulting from programming errors in your SQL code have a severity level in the range 11-16. Severity levels 17-25 indicate resource problems, hardware problems or internal problems in SQL Server, and if the severity is 20 or higher, the connection is terminated. For the long story, see the section More on Severity Levels for some interesting tidbits. For system messages you can find the severity level in master..sysmessages, but for some messages SQL Server employs a different severity level than what's in sysmessages.

State – a value between 0 and 127. The meaning of this item is specific to the error message, but Microsoft has not documented these values, so this value is rarely of interest to you.

Procedure – in which stored procedure, trigger or user-defined function the error occurred. Blank if the error occurred in a plain batch of SQL statements (including dynamic SQL).

Line – Line number within the procedure/function/trigger/batch the error occurred. A line number of 0 indicates that the problem occurred when the procedure was invoked.

Message text – the actual text of the message that tells you what went wrong. You can find this text in master..sysmessages, or rather a template for it, with placeholders for names of databases, tables etc.

As I mentioned the client is responsible for the formatting of the error message, and for messages with a severity level with 10 or lower, most client programs print only the message text, but not severity level, procedure etc. In fact, we see an example of this above. The text The statement has been terminated is a message on its own, message 3621.

When you write your own client program, you can choose your own way to display error messages. You may be somewhat constrained by what your client library supplies to you. The full information is available with low-level interfaces such as DB-Library, ODBC or the OLE DB provider for SQL Server. On the other hand, in ADO you only have access to the error number and the text of the message.

There are two ways an error message can appear: 1) an SQL statement can result in an error (or a warning) 2) you emit it yourself with RAISERROR (or PRINT). Let's take a brief look at RAISERROR here. Here is sample statement:

RAISERROR('This is a test', 16, 1)

Here you supply the message text, the severity level and the state. The output is:

Server: Msg 50000, Level 16, State 1, Line 1
This is a test

Thus, SQL Server supplies the message number 50000, which is the error number you get when you supply a text string to RAISERROR. (There is no procedure name here, since I ran the statement directly from Query Analyzer.) Rather than a string, you could have a supplied a number of 50001 or greater, and SQL Server would have looked up that number in sysmessages to find the message text. You would have stored that message with the system procedure sp_addmessage. (If you just supply a random number, you will get an error message, saying that the message is missing.) Whichever method you use, the message can include placeholders, and you can provide values for these placeholders as parameters to RAISERROR, something I do not cover here. Please refer to Books Online for details.

As I mentioned State is rarely of interest. With RAISERROR, you can use it as you wish. If you raise the same message in several places, you can provide different values to State so that you can conclude which RAISERROR statement that fired. The command-line tools OSQL and ISQL have a special handling of state: if you use a state of 127, the two tools abort and set the DOS variable ERRORLEVEL to the message number. This can be handy in installation scripts if you want to abort the script if you detect some serious condition. (For instance, that database is not on the level that the installation script is written for.) This behaviour is entirely client-dependent; for instance, Query Analyzer does not react on state 127.

How to Detect an Error in T-SQL – @@error

After each statement in T-SQL, with one single exception that I cover in the next section, SQL Server sets the global variable @@error to 0, unless an error occurs, in which case @@error is set to the number of that error. (Note: these days, the SQL Server documentation refers to @@error as a "function". Being an old-timer, I prefer "global variables" for the entities whose names that start with @@.)

More precisely, if SQL Server emits a message with a severity of 11 or higher, @@error will hold the number of that message. And if SQL Server emits a message with a severity level of 10 or lower, SQL Server does not set @@error, and thus you cannot tell from T-SQL that the message was produced.

But the message number is also the only field of the error message that you easily can access from T-SQL. A common question on the newsgroups is how to retrieve the text of an error message, and for a long time the answer was "you can't". But Mark Williams pointed out to me a way to do it. However it requires that the user to have sysadmin privileges, so you cannot easily use it in an application. I will return to this topic in the section Retrieving the Text of an Error Message.

There is no way to prevent SQL Server from raising error messages. There is a small set of conditions for which you can use SET commands to control whether these conditions are errors or not. We will look closer at these possibilities later, but I repeat that this is a small set, and there is no general way in T-SQL to suppress error messages. You will need to take care of that in your client code. (Another common question on the newsgroups.)

As I mentioned, @@error is set after each statement. Therefore, you should always save the save the value of @@error into a local variable, before you do anything with it. Here is an example of what happens if you don't:

CREATE TABLE notnull(a int NOT NULL)
DECLARE @value int
INSERT notnull VALUES (@value)
IF @@error <> 0
   PRINT '@@error is ' + ltrim(str(@@error)) + '.'

The output is:

Server: Msg 515, Level 16, State 2, Line 3
Cannot insert the value NULL into column 'a', table
'tempdb.dbo.notnull'; column does not allow nulls. INSERT fails.
The statement has been terminated.
@@error is 0.

Here is the correct way.

CREATE TABLE notnull(a int NOT NULL)
DECLARE @err int,
        @value int
INSERT notnull VALUES (@value)
SELECT @err = @@error
IF @err <> 0
   PRINT '@err is ' + ltrim(str(@err)) + '.'
The output is:
Server: Msg 515, Level 16, State 2, Line 3
Cannot insert the value NULL into column 'a', table
'tempdb.dbo.notnull'; column does not allow nulls. INSERT fails.
The statement has been terminated.
@err is 515.

Return Values from Stored Procedures

All stored procedures have a return value, determined by the RETURN statement. The RETURN statement takes one optional argument, which should be a numeric value. If you say RETURN without providing a value, the return value is 0 if there is no error during execution. If an error occurs during execution of the procedure, the return value may be 0, or it may be a negative number. The same is true if there is no RETURN statement at all in the procedure: the return value may be a negative number or it may be 0.

Whether these negative numbers have any meaning, is a bit difficult to tell. It used to be the case, that the return values -1 to -99 were reserved for system-generated return values, and Books Online for earlier versions of SQL Server specified meanings for values -1 to -14. However, Books Online for SQL 2000 is silent on any such reservations, and does not explain what -1 to -14 would mean.

With some occasional exception, the system stored procedures that Microsoft ships with SQL Server return 0 to indicate success and any non-zero value indicates failure.

While there is no law that requires you to follow the same convention for your stored procedures, my strong recommendation is that you use return values solely to indicate success/failure. If you want to return data such as the id for an inserted row, number of affected rows or whatever, use an OUTPUT parameter instead. It follows from the fact that a blank RETURN may return 0, even if there has been an error during execution, that you should be careful to return an explict value yourself if an error occurs in the procedure.

There is one situation when a stored procedure does not return any value at all, leaving the variable receiving the return value unaffected. This is when the procedure is aborted because of a scope-aborting error. We will look more into this later. There is also one situation when the return value is NULL: this happens with remote procedures and occurs when the batch is aborted on the remote server. (Batch-abortion is also something we will look into more later on.)

There is one curious exception to the rule that @@error is set after each statement: a RETURN without parameters does not change the value of @@error, but leaves the variable unchanged. In my opinion, this is not really practically useful. (I owe this information to a correspondent who gave me this tip by e-mail. Alas, I lost his mail due to problems at my ISP, so I can credit him by name.)

@@rowcount

@@rowcount is a global variable reports the number of affected rows in the most recently executed statement. Just like @@error you need to save it in a local variable if you want to use the value later, since @@rowcount is set after each statement. Since with SET you can only assign variable at a time, you must use SELECT if you need to save both @@error and @@rowcount into local variables:

SELECT @err = @@error, @rowc = @@rowcount

(For this reason, I prefer to always use SELECT for variable assignment, despite Microsoft's recommendations to use SET.)

In T-SQL it is not an error if, for instance, an UPDATE statement did not affect any rows. But it can of course indicate an error in your application, as it could be an error if a SELECT returns more that one row. For these situations, you can check @@rowcount and raise an error and set a return value, if @@rowcount is not the expected value.

@@trancount

@@trancount is a global variable which reflects the level of nested transactions. Each BEGIN TRANSACTION increases @@trancount by 1, and each COMMIT TRANSACTION decreases @@trancount by 1. Nothing is actually committed until @@trancount reaches 0. ROLLBACK TRANSACTION rolls back everything to the outermost BEGIN TRANSACTION (unless you have used the fairly exotic SAVE TRANSACTION), and forces @@trancount to 0, regards of the previous value.

When you exit a stored procedure, if @@trancount does not have the same value as it had when the procedure commenced execution, SQL Server raises error 266. This error is not raised, though, if the procedure is called from a trigger, directly or indirectly. Neither is it raised if you are running with SET IMPLICIT TRANSACTIONS ON.

More on Severity Levels

In this section we will look a little closer on the various severity levels.

0 Messages with Level 0 are purely informational. A PRINT statement produces a message on severity level 0. These messages do not set @@error. Most query tools prints only the text part of a level 0 message.
1-9 These levels, too, are for informational messages/warnings. I cannot recall that I have encountered this from SQL Server, but I've used it myself in RAISERROR at times. Query Analyzer and SQL Management Studio prints the message number, the level and the state, but not the procedure and line number for these messages.
10 This level does not really exist. It appears that SQL Server internally converts level 10 to level 0, both for its own messages when you use level 10 in RAISERROR.
11-16 These levels indicate a regular programming error of some sort. But it is not the case that level 16 is more serious than level 11. Rather it appears to be a somewhat random categorisation. Books Online gives no details on what the levels might mean, but SQL Server MVP Jacco Schalkwijk pointed out to me that there is a drop-down box in the dialog for defining alerts in Enterprise Manager and SQL Management Studio which has description on these levels. Here is what the drop-down box has to say:

11 – Specified Database Object Not Found
12 – Unused
13 – User Transaction Syntax Error
14 – Insufficient Permission
15 – Syntax Error in SQL Statements
16 – Miscellaneous User Error

My experience is that it may not always be this way, but there certain are matches. Deadlock, for instance is level 13. (So now you know what a User Transaction Syntax Error is!)

17-25 Messages with any of these severity levels indicate some sort of resource problem (for instance running out of disk space), or internal error in SQL Server, or a problem with the operating system or hardware. The higher the severity, the more serious problems. These levels are documented in in the setion Troubleshooting->Error Messages->Error Message Formats->Error Message Severity Levels in Books Online.
19-25 To use level 19 or higher in RAISERROR you must use the WITH LOG option, and you must have sysadmin rights.
20-25 Errors with these severity levels are so fatal, that they always terminate the connection.

What Happens when an Error Occurs?

Many programming languages have a fairly consistent behaviour when there is a run-time error. Common is that the execution simply terminates in case of an error, unless you have set up an exception handler that takes care the error. In other languages, some error variable is set and you have to check this variable. T-SQL is confusing, because depending on what error that occurs and in which context it occurs, SQL Server can take no less than four different actions. I first give an overview of these alternatives, followed by a more detailed discussion of which errors that cause which actions. I then discuss two special cases: trigger context and user-defined functions.

The Possible Actions

These are the four main possible actions SQL Server can take:

Statement-termination. The current statement is aborted and rolled back. Execution continues on the next statement. Any open transaction is not rolled back. @@error is set to the number of the error. Since the statement is rolled back, this means that if you run an UPDATE statement that affects 1000 rows, and for one row a CHECK constraint is violated, none of the rows will be updated. But if the UPDATE statement was part of a longer transaction, the effect of the preceding INSERT, UPDATE or DELETE statements are not affected. You need to issue a ROLLBACK TRANSACTION yourself to undo them.

Scope-abortion. The current scope (stored procedure, user-defined function, or block of loose SQL statements, including dynamic SQL) is aborted, and execution continues on the next statement in the calling scope. That is, if stored procedure A calls B and B runs into a scope-aborting error, execution continues in A, just after the call to B. @@error is set, but the aborted procedure does not have a return value, but the variable to receive the return value is unaffected. As for statement-termination, any outstanding transaction is not affected, not even if it was started by the aborted procedure.

Batch-abortion. The execution of the entire batch – that is, the block of SQL statements that the client submitted to SQL Server – is aborted. Any open transaction is rolled back. @@error is still set, so if you would retrieve @@error first in the next batch, you would see a non-zero value. There is no way you can intercept batch-abortion in T-SQL code. (Almost. We will look a possibility using linked servers later on.)

Connection-termination. The client is disconnected and any open transaction is rolled back. In this case there is no @@error to access.

One can note from this, that there are two things that cannot happen:

But I like to stress that this is based on my own observations. I have found no documentation that actually states that these two cases cannot occur under any circumstances.

The above caters for most of the error situations in SQL Server, but since a hallmark of the error handling in SQL Server is inconsistency, every now and then I discover some new odd situation. I am overlooking these cases here, not to burden the reader with too many nitty-gritty details.

There is however, one more situation you should be aware of and that is batch-cancellation. The client may at any time tell SQL Server to stop executing the batch, and SQL Server will comply more or less immediately. In this situation SQL Server will not roll back any open transaction. (In the general case that is. It seems that if the T-SQL execution is in a trigger, when the cancellation request comes, then there is a rollback.) However, if the current statement when the cancellation request comes in is an UPDATE, INSERT or DELETE statement, then SQL Server will roll back the updates from that particular statement. Batch-cancellation may occur because an explicit call to a cancellation method in the client code, but the most common reason is that a query timeout in the client library expires. ODBC, OLE DB, ADO and ADO.Net all have a default timeout of 30 seconds. (Which judging from the questions on the newsgroups, many programmers believe to come from SQL Server, but not so.)

When Does SQL Server Take which Action?

As you may guess, it depends on the error which action SQL Server takes, but not only. Context also matters. One is the setting of the command SET XACT_ABORT, which we shall look at in a later section. A special case is trigger context, in which almost all errors abort the batch and this will be the topic for the next section. Right now we will discuss the default context, that is outside triggers and when the setting XACT_ABORT is OFF.

You may guess that the more severe the error is, the more drastic action SQL Server takes, but this is only really true for connection-termination. When it comes to scope-abortion, this occurs for a fairly well-defined family, but I am not sure that I agree with that these errors are less severe than the errors that abort the batch. And there is not really any clear distinction between the errors that abort the batch on the one hand, and those that merely terminate the statement on the other. For this reason, I will first cover connection-termination, then scope-abortion and then the other two together.

Connection-termination

When SQL Server terminates the connection, this is because something really bad happened. The most common reason is an execution error in the SQL Server process itself, e.g. an access violation (that is, attempt to access an illegal memory address), a stack overflow, or an assertion error (a programmer-added check for a certain condition that must be true for his code to work). It could also be a protocol error in the communication between the client library and SQL Server. These errors are normally due to bugs in SQL Server or in the client library, but they can also appear due to hardware problems, network problems, database corruption or severe resource problems.

SQL Server terminates the connection, because it would not be safe to continue execution, as internal process structures may be damaged. In some cases, not only is your connection terminated, but SQL Server as such crashes.

Connection-termination can sometimes be due to errors in your application in so far that you may have written some bad SQL that SQL Server could not cope with. But in such case it is still an SQL Server bug if the connection terminates, because you should get a proper error message. (The error messages in conjunction with connection-termination are often very opaque.)

There is one case, though, where a bug in application code can cause connection-termination on its own, and that is if you have your written your own extended stored procedures or your own OLE objects that you call through the sp_OAxxxxx procedures. An unhandled execution error in such code will terminate your connection and may crash SQL Server as well.

There is one way to terminate the connection from T-SQL: if you issue a RAISERROR statement with a severity level >= 20. To do this you must provide WITH LOG, and you must be sysadmin. Since errors with severities >= 19 may trigger an operator alert, and eventually may alert someone's pager, don't do this just for fun.

Scope-abortion

This appears to be confined to compilation errors. At least I have not seen it happen with any other sort of error. Due to the feature known as deferred name resolution (in my opinion this is a misfeature), compilation errors can happen during run-time too. Consider this example (you can run it in the Northwind database):

CREATE PROCEDURE inner_sp @productid int AS

CREATE TABLE #temp (orderid   int      NOT NULL,
                    orderdate datetime NOT NULL)

PRINT 'This prints.'
BEGIN TRANSACTION

INSERT #temp (orderid, orderdate)
   SELECT o.OrderID, o.OrderDate
   FROM   Orders
   WHERE  EXISTS (SELECT *
                  FROM   [Order Details] od
                  WHERE  od.OrderID = o.OrderID
                    AND  od.ProductID = @productid)

COMMIT TRANSACTION
PRINT 'This does not print.'
go
CREATE PROCEDURE outer_sp AS

DECLARE @ret int
SET @ret = 4711
EXEC @ret = inner_sp 76

PRINT '@@error is ' + ltrim(str(@@error)) + '.'
PRINT '@@trancount is ' + ltrim(str(@@trancount)) + '.'
PRINT '@ret ' + coalesce(ltrim(str(@ret)), 'NULL') + '.'
IF @@trancount > 0 ROLLBACK TRANSACTION
go
EXEC outer_sp
go

Because the table #temp does not exist when you create inner_sp, SQL Server defers examination of the entire INSERT-SELECT statement until run-time. Again, when you invoke inner_sp, SQL Server cannot find #temp and defers building a query plan for the INSERT-SELECT statement until it actually comes to execute the statement. It is first at this point, that SQL Server discovers that the SELECT statement is incorrect (the alias for Orders is missing). And at that precise point, the execution of inner_sp is aborted. Here is the output:

This prints.
Server: Msg 266, Level 16, State 2, Procedure inner_sp, Line 18
Transaction count after EXECUTE indicates that a COMMIT or ROLLBACK TRANSACTION
statement is missing. Previous count = 0, current count = 1.
Server: Msg 107, Level 16, State 1, Procedure inner_sp, Line 9
The column prefix 'o' does not match with a table name or alias name used in the query.
Server: Msg 107, Level 16, State 1, Procedure inner_sp, Line 9
The column prefix 'o' does not match with a table name or alias name used in the query.
Server: Msg 107, Level 16, State 1, Procedure inner_sp, Line 9
The column prefix 'o' does not match with a table name or alias name used in the query.
@@error is 266.
@@trancount is 1.
@ret is 4711.

Note the next-to-last line in the output: inner_sp started a transaction. But just because inner_sp was aborted does not mean that the transaction was rolled back. When you implement you error handling, this is something you need to consider, and I look closer at this in the accompanying article on error handling.

Also observe that @ret never was set, but retained the value it had prior to the call.

Not all compilation errors passes unnoticed when SQL Server loads a procedure. A pure syntax error like a missing parenthesis will be reported when you try to create the procedure. But the list of errors not detected because of deferred name resolution is longer than you might expect. After all, one would expect SQL Server be able to detect the missing alias even if #temp is missing. With some effort, it could even detect the missing alias with the Orders table missing, couldn't it?

Actually, I can offer a way to avoid this problem altogether. My toolset AbaPerls, offerde as freeware that includes a load tool, ABASQL. Before creating a procedure, ABASQL extracts all temp tables in the procedure and creates them, so that SQL Server will flag errors such as missing aliases or columns. ABASQL also checks the SQL code for references to non-existing tables.

Statement-termination and Batch-abortion

These two groups comprise regular run-time errors, such as duplicates in unique indexes, running out of disk space etc. As I have already have discussed, which error that causes which action is not always easy to predict beforehand. This table lists some common errors, and whether they abort the current statement or the entire batch.

Error Aborts
Duplicate primary key. Statement
NOT NULL violation. Statement
Violation of CHECK or FOREIGN KEY constraint. Statement
Most conversion errors, for instance conversion of non-numeric string to a numeric value. BATCH
Attempt to execute non-existing stored procedure. Statement
Missing or superfluous parameter to stored procedure to a procedure with parameters. Statement
Superfluous parameter to a parameterless stored procedure. BATCH
Exceeding the maximum nesting-level of stored procedures, triggers and functions. BATCH
Being selected as a deadlock victim. BATCH
Permission denied to table or stored procedure. Statement
ROLLBACK or COMMIT without any active transaction. Statement
Mismatch in number of columns in INSERT-EXEC. BATCH
Declaration of an existing cursor Statement
Column mismatch between cursor declaration and FETCH statement. Statement.
Running out of space for data file or transaction log. BATCH

I am only able to make out a semi-consistency. Some real fatal errors after which I would not really be interested in continuing execution do abort the batch. The examples here are deadlock victim and running out of disk space. But why would it be more severe to pass a superfluous parameter to a parameterless one, than to one that has parameters? And conversion errors? Are they more severe than a constraint violation? And why not all conversion errors? (We will return to conversion errors, as well as arithmetic errors that I purposely excluded from this table, when we discuss the SET commands ANSI_WARNINGS and ARITHABORT. They belong to the small et of errors, where you have some sort of a choice.)

And don't look to severity levels for help. As noteed above, the severity levels 11-16 is another classification, that don't reflect any difference in severity. Most of the errors above have severity level 16, but being a deadlock victim has severity level 13. (Running out of a disk space, which is a resource problem, is level 17.)

Trigger Context

You have trigger context when you are in a trigger, or you are in a stored procedure, user-defined function or block of dynamic SQL that has been called directly or indirectly from a trigger. That is, somewhere on the call stack, there is a trigger. If you are in trigger context, all errors terminate the batch and roll back the transaction on the spot. (Connection-terminating errors still terminate the connection, of course.)

Well, almost. When it comes to error handling in SQL Server, no rule is valid without an exception. Errors you raise yourself with RAISERROR do not abort the batch, not even in trigger context. Neither does error 266, Transaction count after EXECUTE indicates that a COMMIT or ROLLBACK TRANSACTION statement is missing. This error is simply not raised at all when this condition occurs in trigger context. No, this is not a bug, but it is documented in Books Online, and according to Books Online, error 266 is informational only. (Now, taste that concept: an informational error.)

There is one more way that a trigger can terminate the batch. This happens if @@trancount is 0 when the trigger exits. A trigger always executes in the context of a transaction, since even if there is no multi-statement transaction in progress each INSERT, UPDATE and DELETE statement is its own transaction in SQL Server, and the trigger is part of that transaction. Thus, @@trancount is at least 1 when you enter a trigger, and if it is 0 on exit this means that somewhere has been a ROLLBACK statement. (Or sufficiently many COMMIT to bring @@trancount to 0.) Why this would have to abort the batch? Because the sky is blue. Seriously, I don't know, but it has always been that way, and there is no way you can change it.

The normal use for this is that if you have an integrity check in a trigger you raise a message and roll back the transaction, as in this example.

IF EXISTS (SELECT *
           FROM   inserted i
           JOIN   abainstallhistory inh ON i.inhid = inh.inhid
           WHERE  inh.ss_label <> i.ss_label
              OR  inh.ss_label IS NULL AND i.ss_label IS NOT NULL
              OR  inh.ss_label IS NOT NULL AND i.ss_label IS NULL)
BEGIN
   ROLLBACK TRANSACTION
   RAISERROR('Values on ss_label does not match abainstallhistory.', 16, 1)
   RETURN
END

Thus, this trigger aborts the batch, not because of the RAISERROR, but because of the ROLLBACK TRANSACTION, and since the trigger is permitted to execute to end, the RAISERROR statement is executed.

Errors in User-Defined Functions

User-defined functions are usually invoked as part of a SET, SELECT, INSERT, UPDATE or DELETE statement. What I have found is that if an error appears in a multi-statement table-valued function or in a scalar function, the execution of the function is aborted immediately, and so is the statement the function is part of. Execution continues on the next line, unless the error aborted the batch. In either case, @@error is 0. Thus, there is no way to detect that an error occurred in a function from T-SQL.

The problem does not appear with inline table-functions, since an inline table-valued function is basically a macro that the query processor pastes into the query.

You can also execute scalar functions with the EXEC statement. In this case, execution continues if an error occurs (unless it is a batch-aborting error). @@error is set, and you can check the value of @@error within the function. It can be problematic to communicate the error to the caller though.

Control Over Error Handling

No, SQL Server does not offer much in this area, but we will look at the few possibilities, of which the most important is SET XACT_ABORT ON.

SET XACT_ABORT

What I have said this far applies to when XACT_ABORT is OFF, which is the default. When you issue SET XACT_ABORT ON, the very most of the statement-terminating errors instead become batch-aborting errors. Thus, if you don't want to litter your T-SQL code with checks on @@error, and if you are not interested in trying to recover from the error or invoke some error-logging routine in T-SQL, but you are content with execution being aborted on first error, then XACT_ABORT is for you.

Beware, though, that even when XACT_ABORT is ON, not all errors terminate the batch. Here are the exceptions I know of:

So at a minimum you still need to check @@error after the execution of a stored procedure or a block of dynamic SQL even if you use XACT_ABORT ON.

ARITHABORT, ARITHIGNORE and ANSI_WARNINGS

These three SET commands give you very fine-grained control for a very small set of errors. When a division by zero or an overflow occurs, there are no less four choices.

ARITHABORT and ARITHIGNORE also control domain errors, such as attempt to take the square root of a negative number. But this is error is not covered by ANSI_WARNINGS, so here you only have three choices.

As for what is an overflow, SQL Server has extended the domain of this error to datetime value in a way which is not really intuitive. Consider these two statements:

select convert(datetime, '2003123')  -- This causes a conversion error
select @@error
go
select convert(datetime, '20031234') -- This causes an overflow
select @@error

Thus, if you have a string which conforms syntactically to some date format, but some element is out of range, this particular form of conversion error only aborts the batch with a certain setting – and in other settings it may not cause an error at all.

ANSI_WARNINGS controls a few more errors and warnings. With ANSI_WARNINGS ON, it is an error to assign a character or binary column a value that exceeds the the maximum length of the column, and this terminates the statement. When ANSI_WARNINGS is OFF, this condition is not an error, but the value is silently truncated. The error is never raised for variable assignment. Also, with ANSI_WARNINGS ON, if an aggregate function such as SUM() or MIN() sees a NULL value, you get a warning message. (Thus it does not set @@error, nor terminate the statement.)

When you use ODBC, OLE DB and Query Analyzer (SQL 2000), ANSI_WARNINGS is ON by default. Since some features (indexed views, index on computed columns and distributed queries) in SQL Server requires ANSI_WARNINGS to be ON, I strongly recommend that you stick to this. Indexed views and index on computed columns also require ARITHABORT to be ON, but I don't think you can rely on it being ON by default.

Finally, I should mention that there is one more SET command in this area: NUMERIC_ROUNDABORT. When ON, the batch is aborted if operation with a decimal data type results in loss of precision. The option is OFF by default, and it must be OFF for indexed views and indexes on computed columns to work.

RAISERROR WITH NOWAIT

SQL Server buffers the output, so an error message or a result set may not appear directly at the client. In many cases, this is not an issue, but if you are running a long-running procedure, you may want to produce diagnostic messages. To have them displayed immediately in the client, you can use the WITH NOWAIT clause to the RAISERROR statement, as in this example:

PRINT 'This message does not display immediately'
WAITFOR DELAY '00:00:05'
RAISERROR ('But this one does', 0, 1) WITH NOWAIT
WAITFOR DELAY '00:00:05'
PRINT 'It''s over now'

Once there is a message with NOWAIT, all that is ahead of the message in the buffer is also passed to the client.

Unfortunately, there is a bug in SQL Server with NOWAIT, which affects you only if you are calling a procedure through RPC (remote procedure call), so that it this case, SQL Server buffers the messages nevertheless. RPC is the normal way to call a procedure from an application (at least it should be), but if you are running a script from OSQL or Query Analyzer, this bug does not affect you.

Duplicates

Normally when you try to insert a value that would be a duplicate in a unique index, this is an error and the statement is rolled back. However, the syntax for the CREATE INDEX statement includes the option IGNORE_DUP_KEY. When this option is in effect, duplicates are merely discarded. The statement is not rolled back, and if the INSERT statement compassed several rows, the rows that do not violate the uniqueness of the index are inserted.

According to Books Online, SQL Server issues a warning when ignoring a duplicate row. However, in real life the message has severity level 16, and thus comes across to the client as an error. Nevertheless, SQL Server does not set @@error, and as I noted the statement is not rolled back, this message falls in none of four categories I have presented. Microsoft has acknowledged the incorrect severity level as a bug, so hopefully this will be fixed in some future version of SQL Server.

This option applies to unique indexes only. It is not available for PRIMARY KEY or UNIQUE constraints.

Using Linked Servers

There is no way to switch off batch-abortion on a general level. But there is actually one way to handle the case in T-SQL, and that is through linked servers. If you call a remote stored procedure, and the procedure runs into a batch-aborting error, the batch in the calling server is not aborted. On return to the local server, @@error holds the value of the error that aborted the batch on the remote server, and the return value of the stored procedure is set to NULL. (At least my tests indicate this. Thus, it is not the same case as when a local procedure dies with scope-abortion, when the return value is not set at all.) It goes without saying, that this is a fairly lame workaround that is only applicable in special situations.

Some notes:

Retrieving the Text of an Error Message

There is no supported way to retrieve the full text of an error message in SQL 2000. You can get a text from master.dbo.sysmessages, but then you only get placeholders for interesting things like which constraint that was violated. To get the full text of the error message in a proper way, you need a client to pick it up and log it.

However, Mark Williams pointed out you can retrieve the full mesage text from within T-SQL with help of DBCC OUTPUTBUFFER. To wit, after an error has been raised, the messge text is in the output buffer for the process.

The output from DBCC OUTPUTBUFFER is a single colunm, where each row as a byte number, a list of hex values, and a textual representation of the hex values. Mark made the effort to extract the message from the last part, and was kind to send me a stored procedure he had written. As I looked at the output from DBCC OUTPUTBUFFER, I found a byte that appeared to hold the length of the message, which helped me to improve Mark's procedure.

Much later I was contacted by Paulo Santos who looked even deeper into the output from DBCC OUTPUTBUFFER and he was able to significantly improve the procedure, and dig out not only the error message, but also severity, procedure and that. My testing shows that it is still not perfect. Sometimes one of several messages are dropped, junk characters appear and not all line numbers reported correctly. But it is far better than nothing at all and you should not expect something which relies on undocumented behaviour to be perfect. (Of course, on SQL 2005 you would use TRY-CATCH and call error_message() in your CATCH handler.)

There is a very significant restriction with this trick: to run DBCC OUTPUTBUFFER you need sysadmin rights even to look at your own spid, so you cannot put this in an application that is to be run by plain users.

The prime source for the stored procedure is at Paulo's web site, where you find the code and some background. In case his site is down or unavailable, you can find a copy of his spGET_LastErrorMessage here as well. (But check his site first, as he may have updates). If you are curious in history, you can also look the original showErrorMessage that Mark and I produced.

TRY-CATCH in SQL 2005

Next version of SQL Server, SQL 2005, code-named Yukon, introduces significant improvements to the error handling in SQL Server. Here is a simple example:

BEGIN TRY
   SELECT convert(smallint, '2003121')
END TRY
BEGIN CATCH
   PRINT 'errno: ' + ltrim(str(error_number()))
   PRINT 'errmsg: ' + error_message()
END CATCH

The output is:

errno: 244
errmsg: The conversion of the varchar value '2003121' overflowed
an INT2 column. Use a larger integer column.

The construct is similar to error-handling concepts in languages like C++. If an error occurs in the TRY block, or in a stored procedure called by the TRY block, execution is transferred to the CATCH block. In the CATCH block, you have access to six new functions: error_number(), error_severity(), error_state(), error_message(), error_procedure() and error_line(), that gives you all parts of the message associated with the error. And, yes, error_message(), is the expanded message with the parameters filled in.

If you are in a transaction, and the error occurred is a batch-abortion error, your transaction will be doomed. This means that you cannot commit or perform any more updates within the transaction – you must roll back.

One caveat is that if you catch an error in this way, the client will never see the error, unless you call RAISERROR in the error handler. Unfortunately, you cannot reraise the exact error message, since RAISERROR does not permit you to use error numbers less than 50000.

Client-side Error Handling

The various client libraries from which you can access SQL Server have their quirks too. Some libraries are low-level libraries like DB-Library, ODBC and the SQLOLEDB provider. Others are higher-level libraries that sit on top of one of the low-level libraries, one example is ADO. If the low-level library has some quirk or limitation, the high-level library is likely to inherit that. The high-level library might also add its own quirks and limitations.

I am covering four libraries here: DB-Library, ODBC, ADO and ADO .Net, although the first two I discuss very briefly, since most devleopers today use ADO or ADO .Net.

DB-Library

When it comes to error handling, DB-Library is probably the best in the game. When SQL Server produces a message – be that an error, a warning or just an informational message such as a PRINT statement – DB-Library invokes a callback routine, and in that callback routine you have full access to all parts of the message: error number, severity level, state, procedure, line number and of course the message itself. You can then set some global variable to determine what should happen when you come back from the DB-Library call that caused the error.

Unfortunately, Microsoft stopped developing DB-Library with SQL 6.5, and you have poor or no support for new features in SQL Server with DB-Library. Thus, I cannot but discourage you from using DB-Library.

ODBC

With ODBC, you have to rely on return-status values, and then retrieve the error message yourself. Exactly how, I have to admit that I am bit foggy on at this point. However, you do have access to all parts of the error message, and you get all messages. This is evidenced by the fact that you get all this information in Query Analyzer which connects through ODBC.

ADO

ADO is not that good when it comes to error handling. First, you don't have full access to the error message. You only get the error number and the error text. You do not get the severity level (so you don't know whether really is an error at all), nor do you get state, procedure or line number. You do get something called SQLState, which is a five-letter code, not related to SQL Server but inherited from ODBC. Another problem is that you do far from always get all error messages, as I will detail below.

The basic operation with ADO appears simple: You submit a command to SQL Server and if there is an error in the T-SQL execution, ADO raises an error, and if you have set up an error handler with On Error Goto, this is where you will wind up. (On Error is for Basic-derived languages. In C++ I suppose you can use try-catch, but I have not verified this.) You can retrieve all messages from SQL Server in the Errors collection on the Connection object.

But there are quite some surprises hiding here. One thing that makes ADO complicated, is that there are so many ways that you can submit a command and retrieve the results. Partly, this is due to that ADO permits you to access other data sources than SQL Server, including non-relational ones. Also, as your "command" you can simply provide a table name. Since this text is about error handling with stored procedures in SQL Server, I disregard other possibilities. But even if you want to invoke a stored procedure, there are a whole lot of choices:

And that's not really all.

What errors you see in your client code, depends on which combination of all these parameters you use. I developed a form, from which I could choose between these parameters, and then I played with a fairly stupid stored procedure which depending on input could cause some errors, generate some PRINT messages and produce some results sets. And there was a great difference in what I got back. When I used SQLOLEDB and client-side cursors, I did not get any of my two PRINT messages in my .Errors collection if there were no errors, whereas with SQLOLEDB and server-side cursors I got both messages. With MSDASQL, I got the first PRINT message, but not the second, no matter the cursor location.

If there were error messages, I did not always get all of them, but at least one error was communicated and an error was raised in the VB code. However, there is a gotcha here, or two depending on how you see it. The first gotcha is that if the stored procedure produces one or more recordsets before the error occurs, ADO will not raise an error until you have walked past those preceding recordsets with .NextRecordset. This is not peculiar to ADO, but as far as I know applies to all client libraries, and is how SQL Server pass the information to the client. The only odd thing with ADO is that many programmers do not use .NextRecordset, or even know about it. I have also found that in some situations ADO may raise an error and say that .NextRecordset is not supported for your provider or cursor type.

The second gotcha is that your procedure may have more recordsets than you can imagine. To wit, INSERT, UPDATE and DELETE statements generate recordsets to report the rowcount, unless the setting NOCOUNT is ON.

Another irritating feature with ADO that I found, was that as soon there had been an error in the stored procedure, all subsequent result sets from the stored procedure were discarded. I could still tell from the return value of the stored procedure that execution had continued. I have found no combination where you can get the result sets that were produced after an error.

ADO also takes the freedom to make its own considerations about what is an error. I found that ADO always considers division by zero to be an error, even if both ARITHABORT and ANSI_WARNINGS are OFF. In this case, SQL Server merely produces a warning, but ADO opts to handle this warning as an error. A good thing in my opinion. Of what I have found, this only happens with division by zero; not with arithmetic errors such as overflow.

Above I said that even if I did not get all errors from SQL Server, ADO would raise an error. This is true as long as we are talking about commands you submit yourself. But ADO can submit commands behind your back, and if they result in errors, ADO may not alert you – even if the abort the batch and thereby rollback any outstanding transaction. This ugly situation is described further in KB article 810100.

Finally, a note on the return value and value of output parameters from a stored procedure. They are accessible from ADO, even if there is an error during execution of the stored procedure (as long the error does causes the procedure to terminate execution). If you use a client-side cursor you can normally access them directly after executing the procedure, whereas with a server-side cursor you must first retrieve all rows in all result sets. (Which means that if .NextRecordset is not supported for your cursor, you may not be able to retrieve the return value.) Beware that if you try to retrieve these values too soon, you will not be able to retrieve them even when you have retrieved all rows.

It is not really the topic for this text, but the reader might want to know my recommendation of what to choose from all these possibilities. And I say that you should use the SQLOLEDB provider (note that MSDASQL is the default), client-side cursors (note that server-side cursors is the default), invoke your stored procedures from the Command object, using adCmdStoredProcedure. Not because this is the best for error handling, but this appears to be the best from an overall programming perspective. (If you make these choices you will get a static read-only cursor.)

ADO .Net

Note: this applies to ADO .Net 1.1. Since some behaviour I describe may be due to bugs or design flaws, earlier or later versions of ADO .Net may be different in some points.

To some extent, ADO .Net is much better fitted than ADO to handle errors and informational messages from SQL Server, but unfortunately neither ADO .Net is without shortcomings.

The ADO .Net classes can be divided into two groups. The disconnected classes that are common for all data sources, and the connected classes that are data-source specific, but.derived from a common interface. A group such of connected classes makes up a .Net Data Provider and each provider has its own name space. Three providers can connect to SQL Server: There is SqlClient, which is specific to SQL Server, and there are the OLE DB and ODBC .Net Data Providers that connect to anything for which there is an OLE DB provider or an ODBC driver. I will refer to them here as OleDb and Odbc, as this is how their namespaces are spelled in the .Net Framework.

If the only data source you target is SQL Server, SqlClient is of course the natural choice. As we shall see, however, there are situations where OleDb may be preferrable. There is even the odd case where Odbc is the best choice, but as I will detail later, you do best to avoid Odbc when connecting to SQL Server.

The three data providers have some common characteristics when it comes to handling of errors and messages from SQL Server, but there are also significant differences. I will first cover the common features.

To invoke a stored procedure from ADO .Net, you need a Command object. (SqlCommand, OleDbCommand or OdbcCommand). Normally you specify the CommandType as StoredProcedure and provide the procedure name as the command text, but you can also use the CommandType Text and specify an EXEC statement.

There are four methods that you can use to invoke a stored procedure from ADO .Net, and I list them here in the order you are most likely to use them:

DataAdapter.Fill Fills a DataTable or a DataSet with the data from the stored procedure. The are several overloaded Fill methods, some of which permit you to pass a CommandBehavior to specify that you want key or schema information, or that you want only a single row or a single result set.
ExecuteNonQuery Performs a command that does not return any result set (or if it does, you are not interested in it). One example is a store procedure that updates data.
ExecuteReader Returns a DataReader object, through which you can access the rows as they come from SQL Server. If there are several result sets, you use .NextResult to traverse them. This is the most general method to access data. Also here you can specify CommandBehavior.
ExecuteScalar Use this method to run a command that produces a result set of a single value.

To test the possible variations, I wrote a simple application in VB .Net, from which I could pass an SQL command or a stored procedure, and select which data provider and which call method to use. For most of the tests, I used a procedure that depending on input parameters would produce results sets, informational or error messages, possibly interleaved. What follows is based on my observations when playing with this application.

If an error occurs during execution of a stored procedure, the method you used to invoke the procedure will raise an exception. Thus, you should always call these methods within a Try-Catch block, so that you can handle the error message in some way. In the exception handler you have access to a provider-specific Exception object with an ErrorCollection, that containts information about the error. What information that is available is specific for the provider.

If you are interested in informational messages, that is messages with a severity ≤ 10, you can set up an InfoMessage event handler, which you register with the Connection object. It seems, though, if there are both errors and informational messages, that the informational messages comes with the exception. In the event handler, too, you have access to the ErrorsCollection from where you can retrieve the individual messages.

As long as you stick to Fill, ExecuteNonQuery and ExecuteScalar, your life is very simple, as all data has been retrieved once you come back, and if there is an error you wind up in your exception handler. Thus, in difference to ADO, you don't have to bother about unexpected result sets and all that. If you want the return value of a stored procedure or the value of output parameters, these are available in the Parameters collection. However, the OleDb and Odbc providers normally do not fill in these values, if an error occurs during execution of a stored procedure.

If you use ExecuteReader, there are a few extra precautions. If the stored procedure first produces a result set, and then a message, you must first call .NextResult before you get an exception, or, for an informational message, any InfoMessage event handler is invoked. In difference to ADO, ADO .Net does not produce extra result sets for the rowcount of of INSERT, UPDATE and DELETE statements. However, under some circumstances, errors and messages may give cause to extraneous result sets.

Beware that if .NextResult throws an exception, it does not return a value, so if you have something like:

Do
   ....
   Try
      more_results = reader.NextResult()
   Catch e as Exception
      MsgBox(e.Message)
   End Try
Loop Until Not more_results
more_results retains the value it had before you called .NextResult. (Caveat: I'm not an experienced .Net programmer, but this is my observation.)

To get the return value from a stored procedure and the value of output parameters when you use ExecuteReader, you first have to retrieve all rows and all result sets for these values to be available.

Just like ADO, ADO .Net can sometimes generate commands behind your back; this appears mainly to happen when you use the CommandBehaviors KeyInfo and SchemaOnly. But in difference to ADO, ADO .Net communicates any SQL errors from these extra commands, and throws an exception in this case too.

So far, it may seem that ADO .Net is lot more well-behaving than ADO. To some extent it is, but I will now will procede to the specifics for each data provider, and this mainly deals with their respective shortcomings.

SqlClient

One very nice thing with SqlClient, is that the SqlError class includes all components of an SQL Server message: server, error number, message text, severity level, state, procedure and line number.

Another good thing with SqlClient, is that in difference to the other two providers, you do almost always get the return value and the value of output parameters from a stored procedure, even if there is an error during execution (provided that the error does not terminate the execution of the procedure, of course.).

But there are a couple of bad things too:

OleDb

In an OleDbErrorCollection, you don't have access to all information about the error from SQL Server, but only the message text and the message number.

Notes on OleDb:

Odbc

In an OdbcErrorCollection, you don't have access to all information about the error from SQL Server, but only the message text and the message number.

Odbc has all sorts of problems with errors and informational messages. If there are several informational messages, Odbc may lose control and fail to return data, including providing the return value and the values of output parameters of stored procedures. It does not matter whether you have declared an InfoMessage event handler. If there are error messages, and you try to retrieve data, you may get exceptions from the ODBC SQL Server driver saying Function sequence error or Associated statement not prepared.

If there are error messages before any result sets are produced, Odbc may not throw an exception for the first error message, but only invoke your InfoMessage event handler. And if you don't have one, you will not even notice that there was an error. Under some circumstances more than one error message may be dropped this way.

Some of these problems may go away if you run with SET NOCOUNT ON, but not all. In general therefore, I'll advice against using the Odbc .Net Data Provider to access SQL Server.

Still, there is one situation where Odbc is your sole choice, and that is if you call a stored procedure that first produces an error message and then a result set. The other two providers never return any data in this situation. With Odbc you can do it – but it is a narrow path to follow. You must have SET NOCOUNT ON. If you only have one result set, you can probably use OdbcDataAdapter.Fill. If there are more than one result set, you must use ExecuteReader, and you must specify the CommandBehavior SingleResult (!). You may get an exception about Function Sequence Error at the end, but by then you have retrieved all your data.

Acknowledgements and Feedback

Thanks to Trevor Morris who pointed out the tidbit on IMPLICIT_TRANSACTIONS and error 266, Mark Williams and Paulo Santos who investigated DBCC OUTPUTBUFFER and SQL Server MVP Jacco Schalkwijk who found the definition of the severity levels 11-16.

If you have suggestions for improvements or corrections on contents, language or formatting, please mail me at esquel@sommarskog.se. If you have technical questions that any knowledgeable person could answer, I encourage you to post to any of the newsgroups microsoft.public.sqlserver.programming or comp.databases.ms-sqlserver.

Revision History

2009-11-29 – Added a reference for the new, and still unfinished article for error handling in SQL 2005.

2008-03-31 – Thanks to Paulo Santos, there is now an improved stored procedure for retrieving the error message through DBCC OUTPUTBUFFER. Actually, the best way to get the scoop is to go to his web site directly.

2006-01-21 – Added a section on how to retrieve the text from an error message, and a description of severity levels 11-16.

2004-12-26 – Rewritten the section on TRY-CATCH in SQL 2005, to adapt for the Beta 2 changes.

2004-05-30 – One more case where error 266 is not raised.

2004-02-17 – Added a section on ADO .Net.

2003-12-03 – Rewrote section on the RETURN statement. Added section on TRY-CATCH in Yukon.

Back to my home page.