You should never just log in and start entering "ALTER TABLE" commands to change a production database. The project I'm on has database on every customer site, and so every change to the database is made in two places, a dump file that is used to create a new database on a new customer site, and an update file that is run on every update which checks your current database version number against the highest number in the file, and updates your database in place. So for instance, the last couple of updates:

if [ $VERSION \< '8.0.108' ] ; then
  psql -U cosuser $dbName << EOF8.0.108
    BEGIN TRANSACTION;
    --
    -- Remove foreign key that shouldn't have been there.
    -- PCR:35665
    --
    ALTER TABLE     migratorjobitems
    DROP CONSTRAINT migratorjobitems_destcmaid_fkey;
    -- 
    -- Increment the version
    UPDATE          sys_info
    SET             value = '8.0.108'
    WHERE           key = 'DB VERSION';
    END TRANSACTION;
EOF8.0.108
fi

if [ $VERSION \< '8.0.109' ] ; then
  psql -U cosuser $dbName << EOF8.0.109
    BEGIN TRANSACTION;
    --
    -- I missed a couple of cases when I changed the legacy playlist
    -- from reporting showplaylistidnum to playlistidnum
    --
    ALTER TABLE     featureidrequestkdcs
    DROP CONSTRAINT featureidrequestkdcs_cosfeatureid_fkey;
    ALTER TABLE     featureidrequestkdcs
    ADD CONSTRAINT  featureidrequestkdcs_cosfeatureid_fkey
    FOREIGN KEY     (cosfeatureid)
    REFERENCES      playlist(playlistidnum)
    ON DELETE       CASCADE;
    --
    ALTER TABLE     ticket_system_ids
    DROP CONSTRAINT ticket_system_ids_showplaylistidnum_fkey;
    ALTER TABLE     ticket_system_ids
    RENAME          showplaylistidnum
    TO              playlistidnum;
    ALTER TABLE     ticket_system_ids
    ADD CONSTRAINT  ticket_system_ids_playlistidnum_fkey
    FOREIGN KEY     (playlistidnum)
    REFERENCES      playlist(playlistidnum)
    ON DELETE       CASCADE;
    -- 
    -- Increment the version
    UPDATE          sys_info
    SET             value = '8.0.109'
    WHERE           key = 'DB VERSION';
    END TRANSACTION;
EOF8.0.109
fi

我相信有更好的方法来做到这一点，但到目前为止，它对我来说是有效的。

我使用SchemaBank来版本控制我所有的数据库模式更改:

从第一天开始，我将我的db模式转储导入其中 我开始用网络浏览器改变我的模式设计(因为它们是基于SaaS /云的) 当我想更新我的db服务器时，我从它生成更改(SQL)脚本并应用到db。在Schemabank中，它们要求我在生成更新脚本之前将工作提交为版本。我喜欢这种练习，这样当我需要的时候，我就可以随时追溯。

我们的团队规则是，在没有存储设计工作之前，永远不要直接接触db服务器。但它发生了，有人可能会被诱惑打破规则，为了方便。我们将再次将模式转储导入到模式银行中，并让它执行diff，如果发现不一致，则对某人进行攻击。虽然我们可以从它生成alter脚本，使我们的db和模式设计同步，但我们讨厌这样做。

顺便说一下，它们还允许我们在版本控制树中创建分支，这样我就可以维护一个用于登台，一个用于生产。还有一个用于编码沙盒。

一个非常整洁的基于web的模式设计工具，带有版本控制和变更管理。

我使用ActiveRecord Migrations。这个Ruby gem可以在Rails项目之外使用，并且有适配器可以处理您遇到的大多数数据库。我的建议:如果你能够在Postgres上运行你的项目，你就会得到事务性的模式迁移。这意味着如果迁移只应用了一半，您不会最终得到一个损坏的数据库。

我总是检查我的数据库结构转储到源代码控制。全数据库转储，但我通常只是压缩和存储。

作为一种规则，我们将所有的对象代码(存储过程、视图、触发器、函数等)都保存在源代码控制中，因为这些对象都是代码，而且几乎所有其他答案都同意，代码属于某种形式的版本控制系统。

As for CREATE, DROP, ALTER statements, etc. (DDL), we developed and use BuildMaster to manage the deployment of these scripts such that they can be run once and only once against a target database (whether they fail or not). The general idea is that developers will upload change scripts into the system and when it comes time for deployment, only the change scripts that haven't been run against the target environment's database will be run (this is managed very similarly to Autocracy's answer). The reason for this separation of script types lies in that once you manipulate a table's structure, add an index, etc., you effectively cannot undo that without writing a brand new script, or restoring the database - as opposed to the object code where you can simply drop a view or stored procedure then recreate it.

例如，当您将生产数据库恢复到集成环境中时，可以看到一些好处，系统会自动准确地知道哪些脚本没有运行，并将更改新恢复的数据库的表结构，使其与开发相关。

仅供参考，几天前Dana也提出了这个问题…源代码控制中的存储过程/DB模式