在git版本控制下存储每个级别的数据库更改就像每次提交时推送整个数据库，每次拉取时恢复整个数据库。 如果您的数据库很容易发生重大更改，并且您无法承担丢失它们的代价，那么您可以更新pre_commit和post_merge钩子。 我对我的一个项目也做了同样的事情，你可以在这里找到方向。

没有原子性就无法做到这一点，如果不使用pg_dump或快照文件系统，就无法获得原子性。

我的postgres实例在zfs上，我偶尔会对它进行快照。它几乎是即时和一致的。

我说不要。数据在任何时候都可能改变。相反，您应该只提交代码、模式和表定义(创建数据库和创建表语句)中的数据模型以及单元测试的示例数据。这就是Laravel的工作方式，提交数据库迁移和种子。

采用一个数据库转储，并对其进行版本控制。这样它就是一个平面文本文件。

我个人建议同时保留一个数据转储和一个模式转储。通过使用diff，可以相当容易地看到从一个修订到另一个修订的模式中发生了哪些变化。

如果您正在进行大的更改，那么您应该有一个用于进行新模式更改的辅助数据库，而不会触及旧数据库，因为正如您所说的，您正在进行一个分支。

Irmin (branching + time travel) Flur.ee (immutable + time travel + graph query) XTDB (formerly called 'CruxDB') (time travel + query) TerminusDB (immutable + branching + time travel + Graph Query!) DoltDB (branching + time-travel + SQL query) Quadrable (branching + remote state verification) EdgeDB (no real time travel, but migrations derived by the compiler after schema changes) Migra (diffing for Postgres schemas/data. Auto-generate migration scripts, auto-sync db state) ImmuDB (immutable + time-travel)

我们曾经在一个标准的LAMP配置上运行一个社交网站。我们有一个活动服务器、测试服务器和开发服务器，以及本地开发人员机器。所有这些都使用GIT进行管理。

On each machine, we had the PHP files, but also the MySQL service, and a folder with Images that users would upload. The Live server grew to have some 100K (!) recurrent users, the dump was about 2GB (!), the Image folder was some 50GB (!). By the time that I left, our server was reaching the limit of its CPU, Ram, and most of all, the concurrent net connection limits (We even compiled our own version of network card driver to max out the server 'lol'). We could not (nor should you assume with your website) put 2GB of data and 50GB of images in GIT.

To manage all this under GIT easily, we would ignore the binary folders (the folders containing the Images) by inserting these folder paths into .gitignore. We also had a folder called SQL outside the Apache documentroot path. In that SQL folder, we would put our SQL files from the developers in incremental numberings (001.florianm.sql, 001.johns.sql, 002.florianm.sql, etc). These SQL files were managed by GIT as well. The first sql file would indeed contain a large set of DB schema. We don't add user-data in GIT (eg the records of the users table, or the comments table), but data like configs or topology or other site specific data, was maintained in the sql files (and hence by GIT). Mostly its the developers (who know the code best) that determine what and what is not maintained by GIT with regards to SQL schema and data.

When it got to a release, the administrator logs in onto the dev server, merges the live branch with all developers and needed branches on the dev machine to an update branch, and pushed it to the test server. On the test server, he checks if the updating process for the Live server is still valid, and in quick succession, points all traffic in Apache to a placeholder site, creates a DB dump, points the working directory from 'live' to 'update', executes all new sql files into mysql, and repoints the traffic back to the correct site. When all stakeholders agreed after reviewing the test server, the Administrator did the same thing from Test server to Live server. Afterwards, he merges the live branch on the production server, to the master branch accross all servers, and rebased all live branches. The developers were responsible themselves to rebase their branches, but they generally know what they are doing.

如果测试服务器上有问题，例如。合并有太多冲突，然后代码被恢复(将工作分支指向'live')， SQL文件永远不会执行。在执行sql文件时，这被认为是一个不可逆的操作。如果SQL文件不能正常工作，则使用Dump恢复DB(开发人员被告知，因为提供了测试不佳的SQL文件)。

今天，我们同时维护一个sql-up和sql-down文件夹，它们具有相同的文件名，开发人员必须测试这两个正在升级的sql文件是否可以同样降级。这最终可以用bash脚本来执行，但是如果有人一直监视升级过程，这是个好主意。

虽然不是很好，但还是可以控制的。希望这能让你深入了解一个真实的、实用的、相对高可用性的站点。也许它有点过时，但仍然被遵循。