Schema Evolution Documentation

Introduction

Schema evolution is the function of updating an existing Django generated database schema to a newer/modified version based upon a newer/modified set of Django models, and/or a set of developer written upgrade scripts.

It's important to note that different developers wish to approach schema evolution in different ways. As detailed in the original SchemaEvolution document (and elsewhere), there are four basic categories of developers:

1. users who trust introspection and never want to touch/see SQL (Malcolm)
2. users who mostly trust introspection but want the option of auto-applied upgrades for specific situations (Wash)
3. users who use introspection-generated SQL, but don't trust it (they want it generated at development and stored for use in production - Kaylee)
4. users who hate introspection and just want auto-application of their own scripts (Zoe)

who wish to perform different combinations of the two basic subtasks of schema evolution:

1. generation of SQL via magical introspection
2. storage and auto-application of upgrade SQL

This implementation of schema evolution should satisfy all four groups, while keeping the complexities of the parts you don't use out of sight. Scroll down to the usage sections to see examples of how each developer would approach their jobs.

Downloading / Installing

This functionality is not yet in Django/trunk, but in a separate schema-evolution branch. To download this branch, run the following:

svn co http://code.djangoproject.com/svn/django/branches/schema-evolution/ django_se_src
ln -s `pwd`/django_se_src/django SITE-PACKAGES-DIR/django

Or, if you're currently running Django v0.96, run the following:

cd /<path_to_python_dir>/site-packages/django/
wget http://kered.org/blog/wp-content/uploads/2007/07/django_schema_evolution-v096patch.txt
patch -p1 < django_schema_evolution-v096patch.txt

The last command will produce the following output:

patching file core/management.py
patching file db/backends/mysql/base.py
patching file db/backends/mysql/introspection.py
patching file db/backends/postgresql/base.py
patching file db/backends/postgresql/introspection.py
patching file db/backends/sqlite3/base.py
patching file db/backends/sqlite3/introspection.py
patching file db/models/fields/__init__.py
patching file db/models/options.py

How To Use: Malcolm

For the most part, schema evolution can be performed via introspection, as long as you're not doing anything too radical. If you have an established application the vast majority of changes are either additions or renames (either tables or columns). Or if you're new to SQL, introspection keeps things very simple for you. To use schema evolution as Malcolm just make changes to your models, run syncdb, and you're done. But like all schema changes, it's wise to preview what is going to be run. To do this, run the following:

$ ./manage sqlevolve app_name

This will output to the command line the SQL to be run to bring your database schema up to date with your model structure.

However not everything can be handled through introspection. A small amount of metadata is used in the cases of model or field renames, so that the introspection code can match up the old field to the new field. (therefore preserving your data)

For renaming a column, use an "aka" attribute:

    # this field used to be called pub_date
    publish_date = models.DateTimeField('date published', aka='pub_date')

If you have renamed this twice and still wish to support migration from both older schemas, "aka"s can be tuples:

    # this field used to be called pub_date
    publish_date = models.DateTimeField('date published', aka=('pub_date','other_old_field_name'))

For renaming a model, add an "aka" field to the Meta section:

# the original name for this model was 'Choice'
class Option(models.Model):
    [...]
    class Meta:
        aka = 'Choice'

And after time you make a series of changes, run sqlevolve or syncdb and your schema changes will be either shown to you or applied for you.

For further examples, scroll down to the Introspection Examples section.

How To Use: Wash

Note that most Malcolm developers (likely new developers) will eventually run up against a limitation inherent in introspection. They love their incredibly intuitive tool but it can't do everything. But they don't want to give it up, because it's a great 90% solution. If only they can add a simple script without having to throw away all the convenient candy, past or future.

All Wash has to do is store a little bit of extra metadata. Namely two things:

1. a fingerprint of the known schema
2. an sql script

in the file 'app_name/schema_evolution.py'. (conveniently next to models.py)

This module looks as follows:

# list of all known schema fingerprints, in order
fingerprints = [
    'fv1:1742830097',
    'fv1:907953071',
    # add future fingerprints here
]

# all of your evolution scripts, mapping the from_version and to_version 
# to a list if sql commands
evolutions = {
    # ('from_fingerprint','to_fingerprint'): ['-- some sql'],
    ('fv1:1742830097','fv1:907953071'): [
        '-- some list of sql statements, constituting an upgrade',
        '-- some list of sql statements, constituting an upgrade',
    ],
}

To create this file, he would first fingerprint his schema with the following command:

$ ./manage sqlfingerprint app_name
Notice: Current schema fingerprint for 'app_name' is 'fv1:1742830097' (unrecognized)

He would add this fingerprint to the end of the 'fingerprints' list in the schema_evolution module, and it would become an automatically recognized schema, ripe for the upgrade. And then he would write an upgrade script, placing it in the 'evolutions' dictionary object, mapped against the current fingerprint and some fake/temporary fingerprint ('fv1:xxxxxxxx'). Finally, he would run his script (either manually or via syncdb), re-fingerprint and save it in both the fingerprints list and the 'to_fingerprint' part of the mapping.

Later, when he runs sqlevolve (or syncdb) against his production database, sqlevolve will detect his current schema and attempt an upgrade using the upgrade script, and then verify it. If it succeeds, will continue applying all available upgrade scripts until one either fails or it reaches the latest database schema version. (more technically, syncdb will recursively apply all available scripts...sqlevolve since it simply prints to the console, only prints the next available script)

Note: Manually defined upgrade scripts always are prioritized over introspected scripts. And introspected scripts are never applied recursively.

This way Wash can continue using introspections for the majority of his tasks, only stopping to define fingerprints/scripts on those rare occasions he needs them.

How To Use: Kaylee

Kaylee, like Wash and Malcolm, likes the time-saving features of automatic introspection, but likes much more control over deployments to "her baby". So she typically still uses introspection during development, but never in production. What she does is instead of saving the occasional "hard" migration scripts like Wash, she saves them all. This builds a neat chain of upgrades in her schema_evolution module which are then applied in series. Additionally, she likes the ability to automatically back out changes as well, so she stores revert scripts (also usually automatically generated at development) in the same module.

How To Use: Zoe

Zoe simply doesn't like the whole idea of introspection. She's an expert SQL swinger and never wants to see it generated for her (much less have those ugly "aka" fields buggering up her otherwise pristine models. She simply writes her own SQL scripts and stores them all in her schema_evolution module.

Introspection Examples

The following documentation will take you through several common model changes and show you how Django's schema evolution introspection handles them. Each example provides the pre and post model source code, as well as the SQL output.

Adding / Removing Fields

Model: version 1

    from django.db import models
    
    class Poll(models.Model):
        question = models.CharField(maxlength=200)
        pub_date = models.DateTimeField('date published')
        author = models.CharField(maxlength=200)
        def __str__(self):
            return self.question
        
    class Choice(models.Model):
        poll = models.ForeignKey(Poll)
        choice = models.CharField(maxlength=200)
        votes = models.IntegerField()
        def __str__(self):
            return self.choice

Model: version 2

    from django.db import models
    
    class Poll(models.Model):
        question = models.CharField(maxlength=200)
        pub_date = models.DateTimeField('date published')
        author = models.CharField(maxlength=200)
        def __str__(self):
            return self.question
    
        # new fields
        pub_date2 = models.DateTimeField('date published')

    class Choice(models.Model):
        poll = models.ForeignKey(Poll)
        choice = models.CharField(maxlength=200)
        votes = models.IntegerField()
        def __str__(self):
            return self.choice
    
        # new fields
        votes2 = models.IntegerField()
        hasSomething = models.BooleanField()
        creatorIp = models.IPAddressField()

Output: v1⇒v2

    BEGIN;
    ALTER TABLE `case01_add_field_poll` ADD COLUMN `pub_date2` datetime NOT NULL;
    ALTER TABLE `case01_add_field_choice` ADD COLUMN `votes2` integer NOT NULL;
    ALTER TABLE `case01_add_field_choice` ADD COLUMN `hasSomething` bool NOT NULL;
    ALTER TABLE `case01_add_field_choice` ADD COLUMN `creatorIp` char(15) NOT NULL;
    COMMIT;

Output: v2⇒v1

    -- warning: as the following may cause data loss, it/they must be run manually
    -- ALTER TABLE `case01_add_field_poll` DROP COLUMN `pub_date2`;
    -- end warning
    -- warning: as the following may cause data loss, it/they must be run manually
    -- ALTER TABLE `case01_add_field_choice` DROP COLUMN `votes2`;
    -- ALTER TABLE `case01_add_field_choice` DROP COLUMN `creatorIp`;
    -- ALTER TABLE `case01_add_field_choice` DROP COLUMN `hasSomething`;
    -- end warning

Renaming Fields

Model: version 1

    from django.db import models

    class Poll(models.Model):
        """this model originally had fields named pub_date and the_author.  you can use 
        either a str or a tuple for the aka value.  (tuples are used if you have changed 
        its name more than once)"""
        question = models.CharField(maxlength=200)
        pub_date = models.DateTimeField('date published', aka='publish_date')
        the_author = models.CharField(maxlength=200, aka='the_author')
        def __str__(self):
            return self.question
        
    class Choice(models.Model):
        poll = models.ForeignKey(Poll)
        choice = models.CharField(maxlength=200)
        votes = models.IntegerField(aka='votes')
        def __str__(self):
            return self.choice

Model: version 2

    from django.db import models
    
    class Poll(models.Model):
        """this model originally had fields named pub_date and the_author.  you can use
        either a str or a tuple for the aka value.  (tuples are used if you have changed
        its name more than once)"""
        question = models.CharField(maxlength=200)
        published_date = models.DateTimeField('date published', aka=('pub_date', 'publish_date'))
        author = models.CharField(maxlength=200, aka='the_author')
        def __str__(self):
            return self.question
        
    class Choice(models.Model):
        poll = models.ForeignKey(Poll)
        choice = models.CharField(maxlength=200)
        number_of_votes = models.IntegerField(aka='votes')
        def __str__(self):
            return self.choice

Output: v1⇒v2

    BEGIN;
    ALTER TABLE `case02_rename_field_poll` CHANGE COLUMN `pub_date` `published_date` datetime NOT NULL;
    ALTER TABLE `case02_rename_field_poll` CHANGE COLUMN `the_author` `author` varchar(200) NOT NULL;
    ALTER TABLE `case02_rename_field_choice` CHANGE COLUMN `votes` `number_of_votes` integer NOT NULL;
    COMMIT;

Renaming Models

Model: version 1

    from django.db import models
    
    class Poll(models.Model):
        question = models.CharField(maxlength=200)
        pub_date = models.DateTimeField('date published')
        author = models.CharField(maxlength=200)
        def __str__(self):
            return self.question
        
    class Choice(models.Model):
        "the original name for this model was 'Choice'"
        poll = models.ForeignKey(Poll)
        choice = models.CharField(maxlength=200)
        number_of_votes = models.IntegerField()
        def __str__(self):
            return self.choice
        class Meta:
            aka = ('Choice', 'OtherBadName')

Model: version 2

    from django.db import models
    
    class Poll(models.Model):
        question = models.CharField(maxlength=200)
        pub_date = models.DateTimeField('date published')
        author = models.CharField(maxlength=200)
        def __str__(self):
            return self.question
        
    class Option(models.Model):
        "the original name for this model was 'Choice'"
        poll = models.ForeignKey(Poll)
        choice = models.CharField(maxlength=200)
        # show that field name changes work too
        votes = models.IntegerField(aka='number_of_votes')
        def __str__(self):
            return self.choice
        class Meta:
            aka = ('Choice', 'BadName')

Output: v1⇒v2

    BEGIN;
    ALTER TABLE `case03_rename_model_choice` RENAME TO `case03_rename_model_option`;
    ALTER TABLE `case03_rename_model_option` CHANGE COLUMN `number_of_votes` `votes` integer NOT NULL;
    COMMIT;

Changing Flags

Model: version 1

    from django.db import models
    
    class Poll(models.Model):
        question = models.CharField(maxlength=200)
        pub_date = models.DateTimeField('date published')
        author = models.CharField(maxlength=200)
        def __str__(self):
            return self.question
        
    class Choice(models.Model):
        "the original name for this model was 'Choice'"
        poll = models.ForeignKey(Poll)
        choice = models.CharField(maxlength=200)
        votes = models.IntegerField()
        def __str__(self):
            return self.choice

    class Foo(models.Model):
        GENDER_CHOICES = (
            ('M', 'Male'),
            ('F', 'Female'),
        )
        gender = models.CharField(maxlength=1, choices=GENDER_CHOICES)

Model: version 2

    from django.db import models
    
    class Poll(models.Model):
        question = models.CharField(maxlength=100)
        pub_date = models.DateTimeField('date published')
        author = models.CharField(maxlength=200)
        def __str__(self):
            return self.question
        
    class Choice(models.Model):
        "the original name for this model was 'Choice'"
        poll = models.ForeignKey(Poll)
        # make sure aka still works with a flag change
        option = models.CharField(maxlength=400, aka='choice')
        votes = models.IntegerField()
        votes2 = models.IntegerField() # make sure column adds still work
        def __str__(self):
            return self.choice
    
    class Foo(models.Model):
        GENDER_CHOICES = (
            ('M', 'Male'),
            ('F', 'Female'),
        )
        gender = models.CharField(maxlength=1, choices=GENDER_CHOICES, db_index=True)
        gender2 = models.CharField(maxlength=1, null=True, unique=True)
            

Output: v1⇒v2

    BEGIN;
    ALTER TABLE `case04_change_flag_poll` MODIFY COLUMN `question` varchar(100) NOT NULL;
    ALTER TABLE `case04_change_flag_foo` ADD COLUMN `gender2` varchar(1) NULL UNIQUE;
    ALTER TABLE `case04_change_flag_choice` MODIFY COLUMN `choice` varchar(400) NOT NULL;
    ALTER TABLE `case04_change_flag_choice` CHANGE COLUMN `choice` `option` varchar(400) NOT NULL;
    ALTER TABLE `case04_change_flag_choice` ADD COLUMN `votes2` integer NOT NULL;
    COMMIT;

Criticisms

I _really_ don't like the aka representations in the model. The models file should always be a clean statement of the current state of the model. Migration is about getting an old database to match the currently required models - if I don't have a legacy database, I don't really want the cruft hanging around in my models. Migration plans or historical models really should be kept out of the core model file, IMHO.

We currently store all sorts of non-DB related metadata in the model that arguably should not be there, including presentation information. We do this for clarity and convenience - you would have to duplicate a lot of information otherwise in multiple locations without any obvious direct connection. So which paradigm is more critical here, DRY or MVC? Or do we continue with the status-quo of a common-sense balance? As far as cruft, if you don't have a legacy database, you wouldn't have any aka fields to begin with. And as you phase out legacy support, simply delete them.

Unless I'm missing something, the aka approach doesn't track exactly which name versions correlate with other versions. Consider; your aka chain for one field could indicate a rename from 'foo' to 'bar' to 'foo' to 'bar'; a second field could indicate renames from 'whiz' to 'foo' to 'whiz', etc. A tuple of historical names doesn't tell me which sets of names were in use concurrently; so if I find a database with a field called 'foo' which requires migration, is 'foo' the first field or the second?

Correct, however I thought this to be a highly unlikely scenario, not warranting the extra notational complexity. But just as we support strings and tuples, there is nothing to say we can't extend it to support say a mapping of historical names to date ranges, if the need arises.

The 'aka' approach is ambiguous for all but trivial use cases. It 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. This approach doesn't track which state a given database is in, and doesn't apply changes in blocks appropriate to versioned changes.

It does not matter how you get from v1 => v3, as long as you get there with minimum theoretical information loss. The following:

v1 => v2 => v3

1. v1 t1:{A}
2. add_field(B);
3. add_field(C);
4. del_field(A);
5. rename_table(t1,t2);
6. v2 t2{B,C}
7. add_field(A);
8. del_field(C);
9. v3 t2:{A,B}

is functionally equivalent to:

v1 => v3

1. v1 t1:{A}
2. add_field(B);
3. rename_table(t1,t2);
4. v3 t2:{A,B}

And this can be supported completely through introspection + metadata about what tables and columns used to be called. If you load v2 or v3, the available information can get you there from v1, and if you load v3, the available information can get you there from v1 or v2.

A more detailed breakdown of this critique is available ​here, complete with working code examples.

Future Work

The biggest missing piece I believe to be changing column types. For instance, say you currently have:

    ssn = models.IntegerField()

Which you want to change into:

    ssn = models.CharField(maxlength=12)

Schema evolution should generate SQL to add the new column, push the data from the old to the new column, then delete the old column. Warnings should be provided for completely incompatible types or other loss-of-information scenarios.

The second biggest missing piece is foreign/m2m key support.

Lastly, for the migration scripts, sometimes it's easier to write python than it is to write sql. I intend for you to be able to interleave function calls in with the sql statements and have the schema evolution code just Do The Right Thing(tm). But this isn't coded yet.

Conclusion

That's pretty much it. If you can suggest additional examples or test cases you think would be of value, please email me at public@….