Why?

SQLAlchemy already has a versioning extension. This extension however is very limited. It does not support versioning entire transactions.

Hibernate for Java has Envers, which had nice features but lacks a nice API. Ruby on Rails has papertrail, which has very nice API but lacks the efficiency and feature set of Envers.

As a Python/SQLAlchemy enthusiast I wanted to create a database versioning tool for Python with all the features of Envers and with as intuitive API as papertrail. Also I wanted to make it _fast_ keeping things as close to the database as possible.

Features

Installation

pip install SQLAlchemy-Continuum

Basics

In order to make your models versioned you need two things:

import sqlalchemy as sa
from sqlalchemy_continuum import make_versioned
make_versioned(user_cls=None)
class Article(Base):
    __versioned__ = {}
    __tablename__ = 'article'
    id = sa.Column(sa.Integer, primary_key=True, autoincrement=True)
    name = sa.Column(sa.Unicode(255))
    content = sa.Column(sa.UnicodeText)
# after you have defined all your models, call configure_mappers:
sa.orm.configure_mappers()

After this setup SQLAlchemy-Continuum does the following things:

When the models have been configured either by calling configure_mappers() or by accessing some of them the first time, the following things become available:

from sqlalchemy_continuum import version_class, parent_class
version_class(Article)  # ArticleHistory class
parent_class(version_class(Article))  # Article class

Versions and transactions

At the end of each transaction SQLAlchemy-Continuum gathers all changes together and creates version objects for each changed versioned entity. Continuum also creates one TransactionLog entity and N number of TransactionChanges entities per transaction (here N is the number of affected classes per transaction). TransactionLog and TransactionChanges entities are created for transaction tracking.

article = Article(name=u'Some article')
session.add(article)
session.commit()
article.versions[0].name == u'Some article'
article.name = u'Some updated article'
session.commit()
article.versions[1].name == u'Some updated article'

Operation types

When changing entities and committing results into database Continuum saves the used operations (INSERT, UPDATE or DELETE) into version entities. The operation types are stored by default to a small integer field named 'operation_type'. Class called 'Operation' holds convenient constants for these values as shown below:

from sqlalchemy_continuum import Operation
article = Article(name=u'Some article')
session.add(article)
session.commit()
article.versions[0].operation_type == Operation.INSERT
article.name = u'Some updated article'
session.commit()
article.versions[1].operation_type == Operation.UPDATE
session.delete(article)
session.commit()
article.versions[2].operation_type == Operation.DELETE

Version traversal

first_version = article.versions[0]
first_version.index
# 0
second_version = first_version.next
assert second_version == article.versions[1]
second_version.previous == first_version
# True
second_version.index
# 1

Changeset

Continuum provides easy way for getting the changeset of given version object. Each version contains a changeset property which holds a dict of changed fields in that version.

article = Article(name=u'New article', content=u'Some content')
session.add(article)
session.commit(article)
version = article.versions[0]
version.changeset
# {
#   'id': [None, 1],
#   'name': [None, u'New article'],
#   'content': [None, u'Some content']
# }
article.name = u'Updated article'
session.commit()
version = article.versions[1]
version.changeset
# {
#   'name': [u'New article', u'Updated article'],
# }
session.delete(article)
version = article.versions[1]
version.changeset
# {
#   'id': [1, None]
#   'name': [u'Updated article', None],
#   'content': [u'Some content', None]
# }

SQLAlchemy-Continuum also provides a utility function called changeset. With this function you can easily check the changeset of given object in current transaction.

from sqlalchemy_continuum import changeset
article = Article(name=u'Some article')
changeset(article)
# {'name': [None, u'Some article']}

Version relationships

Each version object reflects all parent object relationships. You can think version object relations as 'relations of parent object in given point in time'.

Lets say you have two models: Article and Category. Each Article has one Category. In the following example we first add article and category objects into database.

Continuum saves new ArticleVersion and CategoryVersion records in the background. After that we update the created article entity to use another category. Continuum creates new version objects accordingly.

Lastly we check the category relations of different article versions.

category = Category(name=u'Some category')
article = Article(
    name=u'Some article',
    category=category
)
session.add(article)
session.commit()
article.category = Category(name=u'Some other category')
session.commit()
article.versions[0].category.name  # u'Some category'
article.versions[1].category.name  # u'Some other category'

The logic how SQLAlchemy-Continuum builds these relationships is within the RelationshipBuilder class.

Relationships to non-versioned classes

Let's take previous example of Articles and Categories. Now consider that only Article model is versioned:

class Article(Base):
    __tablename__ = 'article'
    __versioned__ = {}
    id = sa.Column(sa.Integer, autoincrement=True, primary_key=True)
    name = sa.Column(sa.Unicode(255), nullable=False)
class Category(Base):
    __tablename__ = 'tag'
    id = sa.Column(sa.Integer, autoincrement=True, primary_key=True)
    name = sa.Column(sa.Unicode(255))
    article_id = sa.Column(sa.Integer, sa.ForeignKey(Article.id))
    article = sa.orm.relationship(
        Article,
        backref=sa.orm.backref('categories')
    )

Here Article versions will still reflect the relationships of Article model but they will simply return Category objects instead of CategoryVersion objects:

category = Category(name=u'Some category')
article = Article(
    name=u'Some article',
    category=category
)
session.add(article)
session.commit()
article.category = Category(name=u'Some other category')
session.commit()
version = article.versions[0]
version.category.name                   # u'Some other category'
isinstance(version.category, Category)  # True

Dynamic relationships

If the parent class has a dynamic relationship it will be reflected as a property which returns a query in the associated version class.

class Article(Base):
    __tablename__ = 'article'
    __versioned__ = {}
    id = sa.Column(sa.Integer, autoincrement=True, primary_key=True)
    name = sa.Column(sa.Unicode(255), nullable=False)
class Tag(Base):
    __tablename__ = 'tag'
    __versioned__ = {}
    id = sa.Column(sa.Integer, autoincrement=True, primary_key=True)
    name = sa.Column(sa.Unicode(255))
    article_id = sa.Column(sa.Integer, sa.ForeignKey(Article.id))
    article = sa.orm.relationship(
        Article,
        backref=sa.orm.backref(
            'tags',
            lazy='dynamic'
        )
    )
article = Article()
article.name = u'Some article'
article.content = u'Some content'
session.add(article)
session.commit()
tag_query = article.versions[0].tags
tag_query.all()  # return all tags for given version
tag_query.count()  # return the tag count for given version

Revert update

article = Article(name=u'New article', content=u'Some content')
session.add(article)
session.commit(article)
version = article.versions[0]
article.name = u'Updated article'
session.commit()
version.revert()
session.commit()
article.name
# u'New article'

Revert delete

article = Article(name=u'New article', content=u'Some content')
session.add(article)
session.commit(article)
version = article.versions[0]
session.delete(article)
session.commit()
version.revert()
session.commit()
# article lives again!
session.query(Article).first()

Revert relationships

Sometimes you may have cases where you want to revert an object as well as some of its relation to certain state. Consider the following model definition:

class Article(Base):
    __tablename__ = 'article'
    __versioned__ = {}
    id = sa.Column(sa.Integer, autoincrement=True, primary_key=True)
    name = sa.Column(sa.Unicode(255))
class Tag(Base):
    __tablename__ = 'tag'
    __versioned__ = {}
    id = sa.Column(sa.Integer, autoincrement=True, primary_key=True)
    name = sa.Column(sa.Unicode(255))
    article_id = sa.Column(sa.Integer, sa.ForeignKey(Article.id))
    article = sa.orm.relationship(Article, backref='tags')

Now lets say some user first adds an article with couple of tags:

article = Article(
    name=u'Some article',
    tags=[Tag(u'Good'), Tag(u'Interesting')]
)
session.add(article)
session.commit()

Then lets say another user deletes one of the tags:

tag = session.query(Tag).filter_by(name=u'Interesting')
session.delete(tag)
session.commit()

Now the first user wants to set the article back to its original state. It can be achieved as follows (notice how we use the relations parameter):

article = session.query(Article).get(1)
article.versions[0].revert(relations=['tags'])
session.commit()

How many transactions have been executed?

from sqlalchemy_continuum import transaction_class
Transaction = transaction_class(Article)
Transaction.query.count()

Querying for entities of a class at a given revision

In the following example we find all articles which were affected by transaction 33.

session.query(ArticleVersion).filter_by(transaction_id=33)

Querying for transactions, at which entities of a given class changed

In this example we find all transactions which affected any instance of 'Article' model. This query needs the TransactionChangesPlugin.

TransactionChanges = Article.__versioned__['transaction_changes']
entries = (
    session.query(Transaction)
    .innerjoin(Transaction.changes)
    .filter(
        TransactionChanges.entity_name.in_(['Article'])
    )
)

Querying for versions of entity that modified given property

In the following example we want to find all versions of Article class which changed the attribute 'name'. This example assumes you are using PropertyModTrackerPlugin.

ArticleVersion = version_class(Article)
session.query(ArticleHistory).filter(ArticleVersion.name_mod).all()

Transaction

For each committed transaction SQLAlchemy-Continuum creates a new Transaction record.

Transaction can be queried just like any other sqlalchemy declarative model.

from sqlalchemy_continuum import transaction_class
Transaction = transaction_class(Article)
# find all transactions
session.query(Transaction).all()

UnitOfWork

For each database connection SQLAlchemy-Continuum creates an internal UnitOfWork object. Normally these objects are created at before flush phase of session workflow. However you can also force create unit of work before this phase.

uow = versioning_manager.unit_of_work(session)

Transaction objects are normally created automatically at before flush phase. If you need access to transaction object before the flush phase begins you can do so by calling the create_transaction method of the UnitOfWork class.

transaction = uow.create_transaction(session)

The version objects are normally created during the after flush phase but you can also force create those at any time by calling make_versions method.

uow.make_versions(session)

Workflow internals

Consider the following code snippet where we create a new article.

article = Article()
article.name = u'Some article'
article.content = u'Some content'
session.add(article)
session.commit()

This would execute the following SQL queries (on PostgreSQL)

Usage

For enabling native versioning you need to set native_versioning configuration option as True.

make_versioned(options={'native_versioning': True})

Schema migrations

When making schema migrations (for example adding new columns to version tables) you need to remember to call sync_trigger in order to keep the version trigger up-to-date.

from sqlalchemy_continuum.dialects.postgresql import sync_trigger
sync_trigger(conn, 'article_version')

If you don't use PropertyModTrackerPlugin, then you have to disable it:

sync_trigger(conn, 'article_version', use_property_mod_tracking=False)

Using plugins

from sqlalchemy_continuum.plugins import PropertyModTrackerPlugin
versioning_manager.plugins.append(PropertyModTrackerPlugin())
versioning_manager.plugins  # <PluginCollection [...]>
# You can also remove plugin
del versioning_manager.plugins[0]

Activity

The ActivityPlugin is the most powerful plugin for tracking changes of individual entities. If you use ActivityPlugin you probably don't need to use TransactionChanges nor TransactionMeta plugins.

You can initalize the ActivityPlugin by adding it to versioning manager.

activity_plugin = ActivityPlugin()
make_versioned(plugins=[activity_plugin])

ActivityPlugin uses single database table for tracking activities. This table follows the data structure in activity stream specification, but it comes with a nice twist:

Each Activity has relationships to actor, object and target but it also holds information about the associated transaction and about the last associated transactions with the target and object. This allows each activity to also have object_version and target_version relationships for introspecting what those objects and targets were in given point in time. All these relationship properties use generic relationships of the SQLAlchemy-Utils package.

Limitations

Currently all changes to parent models must be flushed or committed before creating activities. This is due to a fact that there is still no dependency processors for generic relationships. So when you create activities and assign objects / targets for those please remember to flush the session before creating an activity:

article = Article(name=u'Some article')
session.add(article)
session.flush()  # <- IMPORTANT!
first_activity = Activity(verb=u'create', object=article)
session.add(first_activity)
session.commit()

Targets and objects of given activity must have an integer primary key column id.

Create activities

Once your models have been configured you can get the Activity model from the ActivityPlugin class with activity_cls property:

Activity = activity_plugin.activity_cls

Now let's say we have model called Article and Category. Each Article has one Category. Activities should be created along with the changes you make on these models.

article = Article(name=u'Some article')
session.add(article)
session.flush()
first_activity = Activity(verb=u'create', object=article)
session.add(first_activity)
session.commit()

Current transaction gets automatically assigned to activity object:

first_activity.transaction  # Transaction object

Update activities

The object property of the Activity object holds the current object and the object_version holds the object version at the time when the activity was created.

article.name = u'Some article updated!'
session.flush()
second_activity = Activity(verb=u'update', object=article)
session.add(second_activity)
session.commit()
second_activity.object.name  # u'Some article updated!'
first_activity.object.name  # u'Some article updated!'
first_activity.object_version.name  # u'Some article'

Delete activities

The version properties are especially useful for delete activities. Once the activity is fetched from the database the object is no longer available ( since its deleted), hence the only way we could show some information about the object the user deleted is by accessing the object_version property.

session.delete(article)
session.flush()
third_activity = Activity(verb=u'delete', object=article)
session.add(third_activity)
session.commit()
third_activity.object_version.name  # u'Some article updated!'

Local version histories using targets

The target property of the Activity model offers a way of tracking changes of given related object. In the example below we create a new activity when adding a category for article and then mark the article as the target of this activity.

session.add(Category(name=u'Fist category', article=article))
session.flush()
activity = Activity(
    verb=u'create',
    object=category,
    target=article
)
session.add(activity)
session.commit()

Now if we wanted to find all the changes that affected given article we could do so by searching through all the activities where either the object or target is the given article.

import sqlalchemy as sa
activities = session.query(Activity).filter(
    sa.or_(
        Activity.object == article,
        Activity.target == article
    )
)

Flask

FlaskPlugin offers way of integrating Flask framework with SQLAlchemy-Continuum. Flask-Plugin adds two columns for Transaction model, namely user_id and remote_addr.

These columns are automatically populated when transaction object is created. The remote_addr column is populated with the value of the remote address that made current request. The user_id column is populated with the id of the current_user object.

from sqlalchemy_continuum.plugins import FlaskPlugin
from sqlalchemy_continuum import make_versioned
make_versioned(plugins=[FlaskPlugin()])

PropertyModTracker

The PropertyModTrackerPlugin offers a way of efficiently tracking individual property modifications. With PropertyModTrackerPlugin you can make efficient queries such as:

Find all versions of model X where user updated the property A or property B.

Find all versions of model X where user didn't update property A.

PropertyModTrackerPlugin adds separate modified tracking column for each versioned column. So for example if you have versioned model called Article with columns name and content, this plugin would add two additional boolean columns name_mod and content_mod for the version model. When user commits transactions the plugin automatically updates these boolean columns.

TransactionChanges

TransactionChanges provides way of keeping track efficiently which declarative models were changed in given transaction. This can be useful when transactions need to be queried afterwards for problems such as:

The plugin works in two ways. On class instrumentation phase this plugin creates a special transaction model called TransactionChanges. This model is associated with table called transaction_changes, which has only only two fields: transaction_id and entity_name. If for example transaction consisted of saving 5 new User entities and 1 Article entity, two new rows would be inserted into transaction_changes table.

TransactionMeta

TransactionMetaPlugin offers a way of saving key-value data for transations. You can use the plugin in same way as other plugins:

meta_plugin = TransactionMetaPlugin()
versioning_manager.plugins.add(meta_plugin)

TransactionMetaPlugin creates a simple model called TransactionMeta. This class has three columns: transaction_id, key and value. TransactionMeta plugin also creates an association proxy between TransactionMeta and Transaction classes for easy dictionary based access of key-value pairs.

You can easily 'tag' transactions with certain key value pairs by giving these keys and values to the meta property of Transaction class.

from sqlalchemy_continuum import versioning_manager
article = Article()
session.add(article)
uow = versioning_manager.unit_of_work(session)
tx = uow.create_transaction(session)
tx.meta = {u'some_key': u'some value'}
session.commit()
TransactionMeta = meta_plugin.model_class
Transaction = versioning_manager.transaction_cls
# find all transactions with 'article' tags
query = (
    session.query(Transaction)
    .join(Transaction.meta_relation)
    .filter(
        db.and_(
            TransactionMeta.key == 'some_key',
            TransactionMeta.value == 'some value'
        )
    )
)

Global and class level configuration

All Continuum configuration parameters can be set on global level (manager level) and on class level. Setting an option at manager level affects all classes within the scope of the manager's class instrumentation listener (by default all SQLAlchemy declarative models).

In the following example we set 'transaction_column_name' configuration option to False at the manager level.

make_versioned(options={'transaction_column_name': 'my_tx_id'})

As the name suggests class level configuration only applies to given class. Class level configuration can be passed to __versioned__ class attribute.

class User(Base):
    __versioned__ = {
        'transaction_column_name': 'tx_id'
    }

Versioning strategies

Similar to Hibernate Envers SQLAlchemy-Continuum offers two distinct versioning strategies 'validity' and 'subquery'. The default strategy is 'validity'.

Validity

The 'validity' strategy saves two columns in each history table, namely 'transaction_id' and 'end_transaction_id'. The names of these columns can be configured with configuration options transaction_column_name and end_transaction_column_name.

As with 'subquery' strategy for each inserted, updated and deleted entity Continuum creates new version in the history table. However it also updates the end_transaction_id of the previous version to point at the current version. This creates a little bit of overhead during data manipulation.

With 'validity' strategy version traversal is very fast. When accessing previous version Continuum tries to find the version record where the primary keys match and end_transaction_id is the same as the transaction_id of the given version record. When accessing the next version Continuum tries to find the version record where the primary keys match and transaction_id is the same as the end_transaction_id of the given version record.

Subquery

The 'subquery' strategy uses one column in each history table, namely 'transaction_id'. The name of this column can be configured with configuration option transaction_column_name.

After each inserted, updated and deleted entity Continuum creates new version in the history table and sets the 'transaction_id' column to point at the current transaction.

With 'subquery' strategy the version traversal is slow. When accessing previous and next versions of given version object needs correlated subqueries.

Column exclusion and inclusion

With exclude configuration option you can define which entity attributes you want to get versioned. By default Continuum versions all entity attributes.

class User(Base):
    __versioned__ = {
        'exclude': ['picture']
    }
    id = sa.Column(sa.Integer, primary_key=True)
    name = sa.Column(sa.Unicode(255))
    picture = sa.Column(sa.LargeBinary)

Basic configuration options

Here is a full list of configuration options:

Example

class Article(Base):
    __versioned__ = {
        'transaction_column_name': 'tx_id'
    }
    __tablename__ = 'user'
    id = sa.Column(sa.Integer, primary_key=True, autoincrement=True)
    name = sa.Column(sa.Unicode(255))
    content = sa.Column(sa.UnicodeText)

Customizing transaction user class

By default Continuum tries to build a relationship between 'User' class and Transaction class. If you have differently named user class you can simply pass its name to make_versioned:

make_versioned(user_cls='MyUserClass')

If you don't want transactions to contain any user references you can also disable this feature.

make_versioned(user_cls=None)

Customizing versioned mappers

By default SQLAlchemy-Continuum versions all mappers. You can override this behaviour by passing the desired mapper class/object to make_versioned function.

make_versioned(mapper=my_mapper)

Customizing versioned sessions

By default SQLAlchemy-Continuum versions all sessions. You can override this behaviour by passing the desired session class/object to make_versioned function.

make_versioned(session=my_session)

Version tables

By default SQLAlchemy-Continuum creates a version table for each versioned entity table. The version tables are suffixed with '_version'. So for example if you have two versioned tables 'article' and 'category', SQLAlchemy-Continuum would create two version tables 'article_version' and 'category_version'.

By default the version tables contain these columns:

If you are using PropertyModTracker Continuum also creates one boolean field for each versioned field. By default these boolean fields are suffixed with '_mod'.

The primary key of each version table is the combination of parent table's primary key + the transaction_id. This means there can be at most one version table entry for a given entity instance at given transaction.

Transaction tables

By default Continuum creates one transaction table called transaction. Many continuum plugins also create additional tables for efficient transaction storage. If you wish to query efficiently transactions afterwards you should consider using some of these plugins.

The transaction table only contains two fields by default: id and issued_at.

Using vacuum

from sqlalchemy_continuum import vacuum
vacuum(session, User)  # vacuums user version

Schema tools

Troubleshooting

If alembic didn't detect any changes or generates reversed migration (tries to remove *_version tables from database instead of creating), make sure that configure_mappers was called by alembic command.

changeset

from sqlalchemy_continuum import changeset
article = Article(name=u'Some article')
changeset(article)
# {'name': [u'Some article', None]}

count_versions

article = Article(name=u'Some article')
count_versions(article)  # 0
session.add(article)
session.commit()
count_versions(article)  # 1

get_versioning_manager

is_modified

article = Article()
is_modified(article)  # False
article.name = 'Something'
is_modified(article)  # True

is_modified_or_deleted

is_session_modified

is_versioned

is_versioned(Article)  # True
article = Article()
is_versioned(article)  # True

parent_class

parent_class(ArticleVersion)  # Article class

transaction_class

from sqlalchemy_continuum import transaction_class
transaction_class(Article)  # Transaction class

version_class

version_class(Article)  # ArticleVersion class

versioned_objects

version_table