Database Deployment Strategies (SQL Server)

问题:

I am looking for a way to do daily deployments and keep the database scripts in line with releases.

Currently, we have a fairly decent way of deploying our source, we have unit code coverage, continuous integration and rollback procedures.

The problem is keeping the database scripts in line with a release. Everyone seems to try the script out on the test database then run them on live, when the ORM mappings are updated (that is, the changes goes live) then it picks up the new column.

The first problem is that none of the scripts HAVE to be written anywhere, generally everyone "attempts" to put them into a Subversion folder but some of the lazier people just run the script on live and most of the time no one knows who has done what to the database.

The second issue is that we have 4 test databases and they are ALWAYS out of line and the only way to truly line them back up is to do a restore from the live database.

I am a big believer that a process like this needs to be simple, straightforward and easy to use in order to help a developer, not hinder them.

What I am looking for are techniques/ideas that make it EASY for the developer to want to record their database scripts so they can be ran as part of a release procedure. A process that the developer would want to follow.

Any stories, use cases or even a link would helpful.

回答1:

For this very problem I chose to use a migration tool: Migratordotnet.

With migrations (in any tool) you have a simple class used to perform your changes and undo them. Here's an example:

[Migration(62)]
public class _62_add_date_created_column : Migration
{
    public void Up()
    {
       //add it nullable
       Database.AddColumn("Customers", new Column("DateCreated", DateTime) );

       //seed it with data
       Database.Execute("update Customers set DateCreated = getdate()");

       //add not-null constraint
       Database.AddNotNullConstraint("Customers", "DateCreated");
    }

    public void Down()
    {
       Database.RemoveColumn("Customers", "DateCreated");
    }
}

This example shows how you can handle volatile updates, like adding a new not-null column to a table that has existing data. This can be automated easily, and you can easily go up and down between versions.

This has been a really valuable addition to our build, and has streamlined the process immensely.

I posted a comparison of the various migration frameworks in .NET here: http://benscheirman.com/2008/06/net-database-migration-tool-roundup

回答2:

Read K.Scott Allen's series of posts on database versioning.
We built a tool for applying database scripts in a controlled manner based on the techniques he describes and it works well.
This could then be used as part of the continuous integration process with each test database having changes deployed to it when a commit is made to the URL you keep the database upgrade scripts in. I'd suggest having a baseline script and upgrade scripts so that you can always run a sequence of scripts to get a database from it's current version to the new state that is needed.
This does still require some process and discipline from the developers though (all changes need to be rolled into a new version of the base install script and a patch script).

回答3:

We've been using SQL Compare from RedGate for a few years now:

http://www.red-gate.com/products/index.htm

The pro version has a command line interface that you could probably use to setup your deployment procedures.

回答4:

We use a modified version of the database versioning described by K. Scott Allen. We use the Database Publishing Wizard to create the original baseline script. Then a custom C# tool based on SQL SMO to dump the stored procedures, views and user functions. Change scripts which contain schema and data changes are generated by Red Gate tools. So we end up with a structure like

Database\
    ObjectScripts\ - contains stored procs, views and user funcs 1-per file
    \baseline.sql - database snapshot which includes tables and data
    \sc.01.00.0001.sql - incremental change scripts
    \sc.01.00.0002.sql
    \sc.01.00.0003.sql

The custom tool creates the database if necessary, applies the baseline.sql if necessary, adds a SchemaChanges table if necessary and applies the change scripts as necessary based on what's in the SchemaChanges table. That process occurs as part of a nant build script each time we do a deployment build via cc.net.

If anyone wants the source code to the schemachanger app I can throw it up on codeplex/google or wherever.

回答5:

If you are talking about trying to keep database schemas in sync, try using Red Gate SQL Comparison SDK. Build a temp database based on a create script (newDb) - this is what you want your database to look like. Compare newDb against your old database (oldDb). Get a change set from that comparison and apply it using Red Gate. You could build this upgrade process into you tests, and you can try and get all the devs to agree that there is one place where the create script for the database is kept. This same practice works well for upgrading your database across several versions and running data migration scripts and processes between each step (using an XML doc to map the create and data migration scripts)

Edit: With Red Gate technique, you only are concerned with create scripts, not upgrade scripts since Red Gate comes up with the upgrade script. It will also let you drop and create indexes, stored procedures, functions, etc.

回答6:

Go here:

https://blog.codinghorror.com/get-your-database-under-version-control/

Scroll down a bit to the list of 5 links to the odetocode.com website. Fantastic five-part series. I would use that as a starting point to get ideas and figure out a process that will work for your team.

回答7:

You should consider using a build tool like MSBuild or NAnt. We use a combination of CruiseControl.NET, NAnt, and SourceGear Fortress to handle our deployments, including SQL objects. The NAnt db build task calls sqlcmd.exe to update scripts in our dev and staging environments after they're checked into Fortress.

回答8:

We use Visual Studio for Database Professionals and TFS to version and manage our database deployments. This allows us to treat our databases just like code (check out, check in, lock, view version history, branch, build, deploy, test, etc.) and even include them in the same solution files if we wish.

Our developers can work on local databases to avoid stepping on each other's changes in a shared environment. When they check database changes into TFS, we have continuous integration to build, test and deploy to our integrated dev environment. We have separate builds on release branches to create differential deployment scripts for each subsequent environment.

Later, if a bug is discovered in a release, we can go to a release branch and hotfix the code and database at the same time.

This is a great product, but its adoption was hindered early on due to a Microsoft marketing blunder. It was originally a separate product under Team System. This meant in order to use features of the developer edition and database edition at the same time, you were required to step up to the much more expensive Team Suite edition. We (and many other customers) gave Microsoft grief about this, and we were very happy they announced this year that DB Pro has been folded into the developer edition, and that immediately anyone licensed with developer edition can install the database edition.

回答9:

Gus off-handedly mentioned DB Ghost (above) – I second it as a potential solution.

A brief overview of how my company is using DB Ghost:

• After the schema for a new DB has been reasonably settled during initial development, we use the DB Ghost 'Data and Schema Scripter' to create script (.sql) files for all the DB objects (and any static data) and we check-in these script files into source control (the tool separates the objects into folders such as 'Stored Procedures', 'Tables', etc.). At this point, we can use either of the DB GHost 'Packager' or 'Packager Plus' tools to create a stand-alone executable to create a new DB from these scripts.
• All changes to the DB schema are checked-in to source by check-ins to the specific script files.
• At anytime we can use the packager to create an executable to either (a) create a new DB or (b) update an existing DB. Some customization is required for certain path-dependent changes (e.g. changes that require data to be updated), but we have pre-update and post-update scripts that are run.

The 'update' process involves the creation of a clean 'source' DB and then (after pre-update custom scripts), a comparison between the schemas of the source DB and the target DB. DB Ghost updates the target DB to match

We routinely make changes to production DBs (we have 14 customers in 7 different production environments) but inevitably deploy a large-enough set of changes with a DB Ghost update executable (created during our build process). Any production changes that were not checked-in to source (or that were not checked-in to the appropriate branch being released) are LOST. This has forced everyone to check-in changes consistently.

To summarize:

• If you enforce a policy that all DB updates be deployed using a DB Ghost update executable, you can 'force' developers to consistently check-in their changes, regardless of whether they are deployed manually in the interim.
• Adding a step (or steps) to your build process to create a DB Ghost update executable will in-effect perform a test to verify that a DB can be created from scripts (i.e. because DB Ghost creates a 'source' DB, even when creating the update executable package) and if you add a step (or steps) to execute the update package [on any of the four test DBs you mentioned], you can keep your test DBs in line with source.

There are some caveats and some limitations in what changes are 'easily' deployed with this tool (really a suite of related tools), but they are all fairly minor (at least for my company):

• Renaming objects must be done in one of the custom scripts
• The entire DB is always updated (e.g. objects in a single schema can't be updated alone) making it difficult to support customer-specific objects in the main application DB

回答10:

The book Refactoring Databases addresses many of these issues at a conceptual level.

As far as tools go, I know that DB Ghost works well for SQL Server. I have heard that the Data Dude edition of Visual Studio has really been imporved upon in the latest release but I don't have any experience with it.

As far as really pulling off continuous integration style database development, it gets really resource instensive really fast because of the number of database copies you need. It is very doable when the database can fit on a developer workstation but impractical when the database is so large that it needs to be deployed across a grid. To do it you bacically need 1 copy of the database per developer [developers who make DDL changes, not just changes to procs] + 6 common copies. The common copies are as follows:

1. INT DEV --> Developers "check in" their refactoring to INT DEV for integration testing. When integration testing passes, this database is copied over to DEV.
2. DEV --> This is the "official" development copy of the database. INT DEV is refreshed regularly with a copy of DEV. Developers working on new refactorings get a fresh copy of the database from DEV.
3. INT QA --> Same idea as INT DEV except for the QA team. When integration tests pass here, this database is copied over to QA and to DEV*.
4. QA
5. INT PROD --> Same idea as INT QA except for production. When integration tests pass here, this database is copied over to PROD, QA*, and DEV*
6. PROD

*When copying databases across DEV/QA/PROD lines, you will also need to run scripts to update test data relevant to the particular environment (e.g. setting up users in QA that the QA team uses to test but that don't exist in production).

回答11:

One possible solution is to look into implementing DML auditing on your test databases, then just rolling those audit logs into a script for final testing and live deployment. SQL Server 2008 significantly improves on DML auditing, but even SQL Server 2005 supports it via triggers.

回答12:

There are a bunch of links in these posts that I'll want to follow up on (I "rolled my own" system years ago, have to see if there are similarities). One thing you will need, and that I hope is mentioned in these links, is discipline. I don't quite see how any automated system can work if anyone can change anything at any time. (Your question implies that this can happen on your production systems, but obviously that can't be true.)

Having one person (the fabled "database administrator") dedicated to the task of managing changes to databases, particularly production databases, is a very common solution. As for maintaining consistency across X development and testing databases: if it/they are used by many users, once again you are best served by having an individual act as a "clearing house" for changes; if everyone has their own database instance, then they're responsible for keeping it in order, and having a central consistent database "source" will be critical when they need a refreshed baseline database.

Here's a recent Stack Overflow post that may be of interest: how-to-refresh-a-test-instance-of-sql-server-with-production-data-without-using

回答13:

Red Gate has a paper describing how to achieve build automation: http://downloads.red-gate.com/HelpPDF/ContinuousIntegrationForDatabasesUsingRedGateSQLTools.pdf

This is built around SQL Source Control, which integrates with SSMS and your existing source control system.

回答14:

I've written a .NET based tool to handle database versioning in an automated fashion. We have been using this tool in production to handle rolling out database updates (including patches) to multiple environments, keep a log in each database of which scripts have been run, and do it all in an automated fashion. It has a command-line console so you can create batch scripts which use this tool. Check it out: https://github.com/bmontgomery/DatabaseVersioning

回答15:

For what it's worth, this is a real example of a simple, low cost approach used by my former employer (and which I am trying to impress on my current employer as a basic first step).

Add a table called 'DB_VERSION' or similar. In EVERY upgrade script, add a row to that table which can include as little or as many columns as you see fit to describe the upgrade but at a minimum I would suggest { VERSION, EXECUTION_DATE, DESCRIPTION, EXECUTION_USER }. Now you have a concrete record of what has been going on. If someone runs their own unauthorised script you'd still need to follow the advice of the answers above, but this is just a simple way of dramatically improving on your existing versioning control (i.e. none).

Now let's you have an upgrade script from v2.1 to v2.2 of the database and you want to verify the lone maverick guy has actually run it on his database, you can just search for rows where VERSION = 'v2.2' and if you get a result, don't run this upgrade script. Can be built into a console utility app if necessary.