A concern of my schema evolution solution is as follows:

The ‘aka’ approach has some serious flaws. It is ambiguous for all but trivial use cases. It also doesn’t capture the idea that database changes occur in bulk, in sequence. For example, On Monday, I add two fields, remove 1 field, rename a table. That creates v2 of the database. On Tuesday, I bring back the deleted field, and remove one of the added fields, creating v3 of the database. In each stage of the migration, the DB is a stable state; this approach doesn’t track which state a given database is in, and doesn’t apply changes in blocks appropriate to versioned changes.

The fallacy in this is twofold:

1. that an automated introspection/evolution must generate and apply schema changes in the same logical order that a DBA would
2. that keeping intermediate state metadata is always necessary (obviously required from #1)

I argue that the exact path from v1 => v3 is irrelevant, as long as it is functionally equivalent to the DBA generated one, and minimizes information loss. To demonstrate this, I’ve coded the above example into three different models.py files:

from django.db import models

class Russ(models.Model):
    "this model is going to have a bit of a day (v1)"
    a = models.CharField(maxlength=200)

from django.db import models

class WasRuss(models.Model):
    "this model is going to have a bit of a day (v2)"
    b = models.CharField(maxlength=200)
    c = models.CharField(maxlength=200)

    class Meta:
        aka = ('Russ')

from django.db import models

class WasRuss(models.Model):
    "this model is going to have a bit of a day (v3)"
    a = models.CharField(maxlength=200)
    b = models.CharField(maxlength=200)

    class Meta:
        aka = ('Russ')

Now let’s assume we have three users: Alice, Bob and Charles. Alice is the developer and Bob and Charles are sys-admins, deploying her application.

On day one, Alice writes her new model (v1) and calls syncdb to create it as you normally would. She then adds data to the table for testing. But on day two, she decides that her original implementation is inadequate and makes her modifications (v2). But instead of writing and storing her own migration scripts or just tossing all her data, she runs sqlevolve, which gives her the following:

ALTER TABLE `case06_russ_russ` RENAME TO `case06_russ_wasruss`;
ALTER TABLE `case06_russ_wasruss` ADD COLUMN `b` varchar(200) NOT NULL;
ALTER TABLE `case06_russ_wasruss` ADD COLUMN `c` varchar(200) NOT NULL;
-- warning: the following may cause data loss
ALTER TABLE `case06_russ_wasruss` DROP COLUMN `a`;
-- end warning

Now day three rolls around, and she’s changed her model again (v3). Again she run’s sqlevolve to get the following:

ALTER TABLE `case06_russ_wasruss` ADD COLUMN `a` varchar(200) NOT NULL;
-- warning: the following may cause data loss
ALTER TABLE `case06_russ_wasruss` DROP COLUMN `c`;
-- end warning

Which gets her exactly to where she needs to be: a schema identical to what a fresh sqlall would give, without destroying all her data. (she did lose everything in column a, however this is acceptable because an identical loss would come from the versioned scripts she would have written by hand)

Now Bob is a bleeding-edge kind of guy. He likes to stay on top of Alice’s work daily. So, assuming she’s a timely svn commiter, each day he runs the following four commands:

$ /etc/init.d/apache stop
$ svn update
$ ./manage sqlevolve | mysql -u root -p my_db
$ /etc/init.d/apache start

This deploys to his database in two days the exact same two scripts she ran, including the same information loss in column a.

Now Charles is more of a conservative deployer - he only deploys when Alice gives them notice, which happened at the end of days one and three. On day one, his syncdb created the database to v1’s specifications. However on day three, when he runs the same commands Bob ran, the following is deployed to his database:

ALTER TABLE `case06_russ_russ` RENAME TO `case06_russ_wasruss`;
ALTER TABLE `case06_russ_wasruss` ADD COLUMN `b` varchar(200) NOT NULL;

As you can see, it is a different script than either Alice or Bob ran, however it gets him to a functionally equivalent schema, and it gets him there with less data loss. (he gets to keep his column a information)

Now this can be argued as either a wonderful or horrible thing. Should Charles be forced to dump his column a data? In some really huge, highly critical, heavily deployed production environments, maybe. But I have managed such before, and I think those cases are few and far between. Much more likely the user is going to want to keep their data. But if they do, a simple procedural change is all that’s necessary. Alice needs only to dump her generated evolution SQL into versioned migration scripts, ala Mike Heald’s dbmigration tool.

So to wrap up, I hope I’ve demonstrated that the idea of “database changes must occur in bulk, in sequence” is flawed, and that what is key is schema equivalence , not making sure you can recreate the exact same set of scripts at runtime for all users using all versions. But that if you do need to make sure identical scripts are run by all users, this can be easily done still using the evolution functionality through minor procedural changes in development and deployment.

I should also note that all the scripts used in this article were generated with the code already checked into the schema-evolution branch. I encourage you to try it out for yourself. (and send me bug reports if you find them!)

Thanks,
Derek

This entry was posted on Friday, August 3rd, 2007 at 12:44 pm and is filed under programming, open source, django. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.