This plan has several problems and has been modified; new plan is discussed in wiki:RepositoryFormat2

----

One problem with [http://www.rubyonrails.org/ Ruby on Rails'] (very good) schema migration system is the behavior of scripts that depend on outside sources; ie. the application. If those change, there's no guarantee that such scripts will behave as they did before, and you'll get strange results. 

For example, suppose one defines a SQLAlchemy table:
{{{
users = Table('users', metadata, 
    Column('user_id', Integer, primary_key = True),
    Column('user_name', String(16), nullable = False),
    Column('password', String(20), nullable = False)
)
}}}
and creates it in a change script:
{{{
from project import table

def upgrade():
  table.users.create()
}}}

Suppose we later add a column to this table. We write an appropriate change script:
{{{
from project import table

def upgrade():
  # This syntax isn't set in stone yet
  table.users.add_column('email_address', String(60), key='email')
}}}
...and change our application's table definition:
{{{
users = Table('users', metadata, 
    Column('user_id', Integer, primary_key = True),
    Column('user_name', String(16), nullable = False),
    Column('password', String(20), nullable = False),
    Column('email_address', String(60), key='email')  #new column
)
}}}

Modifying the table definition changes how our first script behaves - it will create the table with the new column. This might work if we only apply change scripts to a few database which are always kept up to date (or very close), but we'll run into errors eventually if our migration scripts' behavior isn't consistent.

----

One solution is to generate .sql files from a Python change script at the time it's added to a repository. The sql generated by the script for each database is set in stone at this point; changes to outside files won't affect it.

This limits what change scripts are capable of - we can't write dynamic SQL; ie., we can't do something like this:
{{{
for row in db.execute("select id from table1"):
  db.execute("insert into table2 (table1_id, value) values (:id,42)",**row)
}}}
But SQL is usually powerful enough to where the above is rarely necessary in a migration script:
{{{
db.execute("insert into table2 select id,42 from table1")
}}}
This is a reasonable solution. The limitations aren't serious (everything possible in a traditional .sql script is still possible), and change scripts are much less prone to error.