This is a follow up to #91. Upon further debugging, I have made the observation that instead of translating fizz migrations 1:1 to SQL statements, pop/fizz tries to be smart and bundles together the SQL statements, creating one SQL "file" that has "optimizations" (such as removing alter tables in favor of just creating the table). This of course is very bad because if bugs (such as #91) happen, they actually depend on the existing state of the database. For example, a fresh database is missing the foreign key while a database with migrations partially applied does not.

Another serious issue is that fixing bugs like #91 in fizz will cause divergence of databases because new databases (without any existing state) will properly create the foreign key, while older databases (with the buggy statement) will not. Bringing them into line is basically a huge trial and error/guessing game based on what SQL statements are used when and where and how and what database has already had which migrations applied.

I can not state enough that this is really bad for any system that aims to have reliable SQL migrations - so far that we are definitely going to have to switch to raw SQL and tell our users to level their existing database because there's too much complexity involved in fixing all these nuances.

But there is a path forward in my opinion: Fizz (or pop) should use the fizz templates to generate SQL files for each database, which then should be checked into version control and be applied by the migration logic instead of using fizz on-the-fly. This means that fizz would basically become a CLI command that could be used with e.g. go generate along the lines of:

fizz migrations/*.fizz sql_migration_output/

That would then be a directory along the lines of:

• sql_migration_output
  • 123123_myfizzmigration.mysql.sql
  • 123123_myfizzmigration.postgres.sql
  • 123123_myfizzmigration.sqlite.sql

and so on.

As an intermediate step I would highly encourage to put a big warning in the docs to not use fizz migrations because of these issues. Whatever you do, you will run into this stuff eventually and when you do - and depending on when - you're pretty much fucked, especially if it's another OSS project used by people against many, many databases.