A single logical migration is represented as two separate migration files, one to migrate "up" to the specified version from the previous version, and a second to migrate back "down" to the previous version. These migrations can be provided by any one of the supported migration sources.
The ordering and direction of the migration files is determined by the filenames
used for them. migrate expects the filenames of migrations to have the format:
{version}_{title}.up.{extension}
{version}_{title}.down.{extension}
The title of each migration is unused, and is only for readability. Similarly,
the extension of the migration files is not checked by the library, and should
be an appropriate format for the database in use (.sql for SQL variants, for
instance).
Versions of migrations may be represented as any 64 bit unsigned integer. All migrations are applied upward in order of increasing version number, and downward by decreasing version number.
Common versioning schemes include incrementing integers:
1_initialize_schema.down.sql
1_initialize_schema.up.sql
2_add_table.down.sql
2_add_table.up.sql
...
Or timestamps at an appropriate resolution:
1500360784_initialize_schema.down.sql
1500360784_initialize_schema.up.sql
1500445949_add_table.down.sql
1500445949_add_table.up.sql
...
But any scheme resulting in distinct, incrementing integers as versions is valid.
It is suggested that the version number of corresponding up and down migration
files be equivalent for clarity, but they are allowed to differ so long as the
relative ordering of the migrations is preserved.
The migration files are permitted to be "empty", in the event that a migration is a no-op or is irreversible. It is recommended to still include both migration files by making the whole migration file consist of a comment. If your database does not support comments, then deleting the migration file will also work. Note, an actual empty file (e.g. a 0 byte file) may cause issues with your database since migrate will attempt to run an empty query. In this case, deleting the migration file will also work. For the rational of this behavior see: #244 (comment)
The format of the migration files themselves varies between database systems. Different databases have different semantics around schema changes and when and how they are allowed to occur (for instance, if schema changes can occur within a transaction).
As such, the migrate library has little to no checking around the format of
migration sources. The migration files are generally processed directly by the
drivers as raw operations.
Best practice for writing schema migration is that all migrations should be reversible. It should in theory be possible for run migrations down and back up through any and all versions with the state being fully cleaned and recreated by doing so.
By adhering to this recommended practice, development and deployment of new code is cleaner and easier (cleaning database state for a new feature should be as easy as migrating down to a prior version, and back up to the latest).
As opposed to some other migration libraries, migrate represents up and down
migrations as separate files. This prevents any non-standard file syntax from
being introduced which may result in unintended behavior or errors, depending
on what database is processing the file.
While it is technically possible for an up or down migration to exist on its own without an equivalently versioned counterpart, it is strongly recommended to always include a down migration which cleans up the state of the corresponding up migration.