Avoid Broken Views in SQL Server with Schemabinding
SQL Server views can get really weird when an underlying table is updated. I’ll show some examples and also the one must do for views to avoid problems.
The broken views problem is something I discovered when working on a production system. When I first observed it I was shocked. When I dug deeper into it I was even more shocked at the pure evilness this behaviour could cause. But let’s take it from the start with a simple table of colours and a view on top of it to start with.
CREATE TABLE [dbo].[Colours] ( [Id] INT IDENTITY (1, 1) NOT NULL, [Colour] NVARCHAR (10) NOT NULL, CONSTRAINT [PK_Colours] PRIMARY KEY CLUSTERED ([Id] ASC) ); GO CREATE VIEW [dbo].[BadView] AS SELECT * FROM Colours
Populate the table with some data and test the view.
SELECT * FROM BadView Id Colour ----------- ---------- 1 Red 2 Green 3 Blue 4 White 5 Black
Now, let’s break havoc to the view. I’ll recreate the table with another column
RGB which is inserted right after the
Id column. The
Colour column is renamed to
ColourName. Before running the next test, I’ve inserted the appropriate RGB values in the new column.
SELECT * FROM BadView Id Colour ----------- ------ 1 FF0000 2 00FF00 3 0000FF 4 FFFFFF 5 000000
This is what I consider weird behaviour. The output of the view still contains two columns with the names
That would be good – if it contained the right data. It doesn’t. The
data is taken from the second column of the underlying table. If you’re
not scared enough by this I have some more, even worse things to show…
When misaligned column names and data is discovered it must of course be handled. All problems can be solved in different ways. One way I found in a production system was a select clause like this.
SELECT Id, Colour AS RGB FROM BadView
It works. It produces output with the right column name for the data. My feelings and personal thoughts of that approach are not suitable for printing out.
Id RGB ----------- ------ 1 FF0000 2 00FF00 3 0000FF 4 FFFFFF 5 000000
It was very obvious at this point that the system was brittle and that I would have to be extremely careful during the necessary clean up. Scripting out the existing definition of a view and rerunning it should be safe. Scripting a database object is a way to capture the current definition of it, to be able to restore it. It’s an no-op, right. RIGHT? So that’s what I did (on a backup copy of the DB).
ALTER VIEW [dbo].[BadView] AS SELECT * FROM Colours
Then I once again tested the view.
SELECT * FROM BadView Id RGB ColourName ----------- ------ ---------- 1 FF0000 Red 2 00FF00 Green 3 0000FF Blue 4 FFFFFF White 5 000000 Black
I now was not only scared. I was horrified. The database is not just showing some really weird behaviour. It keeps information about the view, that is not included in the scripted definition of the object. Recreating it with exactly the same script won’t give the same result.
To me this is a horror story. Walking down the road of legacy code on the way to maintenance hell.
Time to look at the cure for this weirdness.
The SCHEMABINDING cure
When I created the
BadView I also created a
GoodView that applied
WITH SCHEMABINDING. When a view is schema bound all objects (tables, functions) that the view is dependent on are locked against modification.
First I tried to create the
GoodView just like
CREATE VIEW [dbo].[GoodView] WITH SCHEMABINDING AS SELECT * FROM Colours
Msg 1054, Level 15, State 6, Procedure GoodView, Line 3 Syntax '*' is not allowed in schema-bound objects.
With schema binding, SQL server requires a more specific and safer select statement.
CREATE VIEW [dbo].[GoodView] WITH SCHEMABINDING AS SELECT Id, Colour FROM dbo.Colours
When doing the rename of the
Colour column to
ColourName the rename operation fails.
Msg 5074, Level 16, State 1, Line 3 The object 'GoodView' is dependent on column 'ColourName'. Msg 4922, Level 16, State 9, Line 3 ALTER TABLE DROP COLUMN ColourName failed because one or more objects access this column.
The message is from the SQL server engine itself. The schemabound view is protected from breaking changes of the underlying table regardless of what client tool is used.
Tooling Support is Required
I found an article: SCHEMABINDING – use or not to use that discusses the consequences of schema binding. I think that the points are valid, but it fails to mention that good tooling support can handle the problems by automatically dropping and recreating any views that are schema bound to a table that is changed. The database deployment support in Visual Studio 2010 and 2012 (much improved in 2012) is an example of a tool that handles this.
What about stored procedures?
So far this post has been about views. Probably more common than views are stored procedures. They at least doesn’t use stale column definitions.
CREATE PROCEDURE [dbo].[MyProcedure] AS SELECT * FROM Colours
Colours table is updated, the stored procedure output is adjusted to include the new column and the new name of the renamed column.
My recommendation is to schema bind all views as the default. If it becomes to hard to change the underlying tables when there are multiple schema bound views a better tool is needed. Dropping the schema binding is not a good solution. I prefer views over stored procedures for simple queries as they can be reused in other queries. There used to be a performance benefit for stored procedures but in modern SQL server versions all queries are compiled and the query plan cached.