Skip to main content

Building deployment scripts

There are specific ways deployment scripts should be structured. We do so to adhere to the following guiding principles:

✅ The script must be idempotent: Always ensure a migration can be run multiple times without causing errors or duplicating data. We never intend to run scripts multiple times on an environment, but the scripts must support it.

✅ The script must avoid breaking changes: Migrations should never delete or rename columns that are still in use by deployed code.

✅ The script must maintain schema integrity: The schema of the database defined in code should map exactly to the schema of the database in all deployed environments.

✅ The script must be backwards compatible: Code should be able to work with both the old and new schema during a rolling deployment. See Database Migrations for how we put this into practice.

Script best practices by database entity​

tip

Migration scripts may not be something that you do often, and it is important that we follow our guidelines to protect against potential downtimes during release or data loss. When writing a script, check each of the entities you're working with against the recommendations below before you introduce a pull request with your changes.

Tables

Creating a table​

When creating a table, you must first check if the table exists:

IF OBJECT_ID('[dbo].[{table_name}]') IS NULL
BEGIN
    CREATE TABLE [dbo].[{table_name}] (
        [Id]                UNIQUEIDENTIFIER NOT NULL,
        ...
        CONSTRAINT [PK_{table_name}] PRIMARY KEY CLUSTERED ([Id] ASC)
    );
END
GO

Deleting a table​

When deleting a table, use IF EXISTS to avoid an error if the table doesn't exist.

DROP IF EXISTS [dbo].[{table_name}]
GO

Columns

Adding a column to a table​

You must first check to see if the column exists before adding it to the table.

IF COL_LENGTH('[dbo].[{table_name}]', '{column_name}') IS NULL
BEGIN
    ALTER TABLE [dbo].[{table_name}]
        ADD [{column_name}] {DATATYPE} {NULL|NOT NULL};
END
GO

Nullability​

When adding a new NOT NULL column to an existing table, please re-evaluate the need for it to truly be required. Do not be afraid of using Nullable<T> primitives in C# and in the application layer, which is almost always going to be better than taking up unnecessary space in the DB per row with a default value, especially for new functionality or features where it will take a very long time to be useful for most row-level data, if at all.

If you do decide to add a NOT NULL column, use a DEFAULT constraint instead of creating the column, updating rows and changing the column. This is especially important for the largest tables like dbo.User and dbo.Cipher. Our version of SQL Server in Azure uses metadata for default constraints. This means we can update the default column value without updating every row in the table (which will use a lot of DB I/O).

This is slow:

IF COL_LENGTH('[dbo].[Table]', 'Column') IS NULL
BEGIN
    ALTER TABLE
        [dbo].[Table]
    ADD
        [Column] INT NULL
END
GO

UPDATE
    [dbo].[Table]
SET
    [Column] = 0
WHERE
    [Column] IS NULL
GO

ALTER TABLE
    [dbo].[Column]
ALTER COLUMN
    [Column] INT NOT NULL
GO

This is better:

IF COL_LENGTH('[dbo].[Table]', 'Column' IS NULL
BEGIN
    ALTER TABLE
        [dbo].[Column]
    ADD
        [Column] INT NOT NULL CONSTRAINT D_Table_Column DEFAULT 0
END
GO

Column order​

When you make corresponding updates to the database schema in code, always add new columns to the end of the column list. Adding columns in between existing ones creates schema disparities between the schema defined in code and the actual column order in our local, staging, and production environments, since the column ADD will always add to the end of the table definition.

This can cause subtle and hard-to-debug issues, particularly when using non-parameterized SQL that relies on implicit column order or performing bulk inserts/updates where tools assume column order consistency.

While well-written code shouldn't depend on column order, some third-party tools and legacy practices might.

Changing a column data type​

You must wrap the ALTER TABLE statement in a conditional block, so that subsequent runs of the script will not modify the data type again.

IF EXISTS (
    SELECT *
    FROM INFORMATION_SCHEMA.COLUMNS
    WHERE COLUMN_NAME = '{column_name}' AND
        DATA_TYPE = '{datatype}' AND
        TABLE_NAME = '{table_name}')
BEGIN
    ALTER TABLE [dbo].[{table_name}]
    ALTER COLUMN [{column_name}] {NEW_TYPE} {NULL|NOT NULL}
END
GO

Adjusting metadata​

When adjusting a table, you should also check to see if that table is referenced in any views. If the underlying table in a view has been modified, you should run sp_refreshview to re-generate the view metadata.

EXECUTE sp_refreshview N'[dbo].[{view_name}]'
GO

Views

Creating or modifying a view​

We recommend using the CREATE OR ALTER syntax for adding or modifying a view.

CREATE OR ALTER VIEW [dbo].[{view_name}]
AS
SELECT
    *
FROM
    [dbo].[{table_name}]
GO

Deleting a view​

When deleting a view, use IF EXISTS to avoid an error if the table doesn't exist.

DROP IF EXISTS [dbo].[{view_name}]
GO

Adjusting metadata​

When altering views, you may also need to refresh modules (stored procedures or functions) that reference that view or function so that SQL Server to update its statistics and compiled references to it.

IF OBJECT_ID('[dbo].[{procedure_or_function}]') IS NOT NULL
BEGIN
    EXECUTE sp_refreshsqlmodule N'[dbo].[{procedure_or_function}]';
END
GO

Functions and stored procedures

Creating or modifying a function or stored procedure​

We recommend using the CREATE OR ALTER syntax for adding or modifying a function or stored procedure.

CREATE OR ALTER {PROCEDURE|FUNCTION} [dbo].[{sproc_or_func_name}]
...
GO

Deleting a function or stored procedure​

When deleting a function or stored procedure, use IF EXISTS to avoid an error if it doesn't exist.

DROP IF EXISTS [dbo].[{sproc_or_func_name}]
GO
warning

When changing a stored procedure, ensure that the corresponding Entity Framework model is updated.

Indices

When creating indexes, especially on heavily used tables, our production database can easily become offline, unusable, hit 100% CPU and many other bad behaviors. It is often best to do this using online index builds so as not to lock the underlying table. This may cause the index operation to take longer, but you will not create an underlying schema table lock which prevents all reads and connections to the table and instead only locks the table of updates during the operation.

A good example is when creating an index on dbo.Cipher or dbo.OrganizationUser, those are heavy-read tables and the locks can cause exceptionally high CPU, wait times and worker exhaustion in Azure SQL.

CREATE NONCLUSTERED INDEX [IX_OrganizationUser_UserIdOrganizationIdStatus]
   ON [dbo].[OrganizationUser]([UserId] ASC, [OrganizationId] ASC, [Status] ASC)
   INCLUDE ([AccessAll])
   WITH (ONLINE = ON); -- ** THIS ENSURES ONLINE **