Imagine: you are writing a Python application of some kind. You have defined a number of complex objects, each with properties of various types. Some of these properties are other objects that are also defined within part of your code. You now need to persist the state of these objects, so that this state can be retrieved in future runs of your program. The Zope Object Database ("ZODB") can help. The ZODB allows you to persist your Python objects easily.
Python's dynamic nature allows developers to quickly develop applications, avoiding the compile cycle and static typing declarations required by other languages. The ZODB offers a similar benefit: developers who use the ZODB can store their objects transparently without any cumbersome mapping of objets to relational database tables. They no longer need to worry about ways of decomposing complex objects in order to fit a relational or filesystem model: using ZODB, programmers can store Python objects in their native, assembled state.
This book will introduce the ZODB in detail and will cover its features step-by-step, using a hands-on approach with many code examples and practical tips.
Looking at the ZODB from 10,000 feet
ZODB takes a minimalist approach, compared to other systems which call themselves "databases". ZODB provides persistence and transaction support, but provides no security, indexing or search facilities. There are third party packages which provide these (and additional) services for the ZODB, but for now let's take a look at its core features.
For a Python developer, the ZODB offers a very familiar environment. Not only does it store native Python objects, but even its internal serialization mechanism is well known to Python programmers: it uses the pickle module included in the Python Standard Library. ZODB is written in Python itself, so in extreme cases developers can look under the hood without needing other tools than the ones they use regularly.
The ZODB is a hierarchical database. There is a root object, initialized when a database is created. The root object is used like a Python dictionary and it can contain other objects (which can be dictionary-like themselves). To store an object in the database, it's enough to assign it to a new key inside its container.
The ZODB has been around for over ten years and has been put into many production environments.
To make an instance of a class automatically persistent, it's enough for its class to inherit from a base "Persistent" class. The ZODB will then take care of saving objects which inherit from this class whenever they are changed. Non-persistent objects can also be saved easily to ZODB, but in some cases it becomes necessary to alert the ZODB when they change.
Transactions are a series of changes to the database that need to be carried out as a unit. That is, either all of the changes in a transaction take place, or none do. Generally, when you are through with a series of changes the transaction is committed and if anything goes wrong it is aborted.
If you have worked with relational databases, transactions should be familiar. Transactional systems need to make sure that the database never gets into an inconsistent state, which they do by supporting four properties, known by the acronym ACID:
- Atomicity. Either all the modifications grouped in a transaction will be written to the database or, if something makes this impossible, the whole transaction will be aborted. This insures that in the event of a write error or a hardware glitch the database will remain in the previous state and avoid inconsistencies.
- Consistency. For writing transactions, these means that no transaction will be allowed if it would leave the database in an inconsistent state. For reading transactions it means that a read operation will see the database in the consistent state it was at the beginning of the transaction, regardless of other transactions taking place at the time.
- Isolation. When changes are made to the database by two different programs, they will not be able to see each other's transactions until they commit their own.
- Durability. This simply means that the data will be safely stored once the transaction is committed. A software or hardware crash will not cause any information to be lost after that.
Since changes made during a single transaction are kept in memory until the transaction is committed, memory usage can skyrocket during a transaction where lots of objects are modified at the same time (say a for loop which changes a property on 100,000 objects). Save points allow developers to commit part of one transaction before it's finished so that changes are written to the database and the memory they occupied is released. This changes in the save point are not committed until the whole transaction is finished, so that if it is aborted, any save points will be rolled back as well.
The ZODB provides a very simple mechanism to roll back any committed transaction. This feature is possible because ZODB keeps track of the database state before and after every transaction. This makes it possible to undo the changes in a transaction, even if more transactions have been committed after it. Of course, if the objects involved in this transaction that we need to undo have changed in later transactions, it will not be possible to undo it because of consistency requirements.
Since every transaction is kept in the database, it's possible to view an object's state as it was in previous transactions and compare it with its current state. This allows a developer to quickly implement simple versioning functionality.
Binary large objects, such as images or office documents, do not need all the versioning facilities that the ZODB offers. In fact, if they were handled as regular object properties, blobs would make the size of a database increase greatly and generally slow things down. That's why the ZODB uses a special storage for blobs, which makes it feasible to easily handle large files up to a few hundred megabytes without performance problems.
Every time an object is read from the database, it's kept in an in-memory LRU cache. Subsequent accesses to this object consume less resources and time. The ZODB manages the cache transparently and automatically pulls out objects which have not been accessed for a long time. The size of the cache can be configured, so that machines with more memory can take better advantage of the feature.
ZODB keeps all versions of the objects stored in it. This means that the database grows with every object modification and it can reach a large size, which may slow it down and consume more space than is necessary. The ZODB allows us to remove old revisions of stored objects via a procedure known as packing. The packing routine is flexible enough to allow only objects older than a specified number of days to be removed, keeping the newer revisions around.
By default, the ZODB stores the database in a single file. The program which manages this is called a file storage. However, the ZODB is built in such a way that other storages can be plugged in without needing to modify its source code. This can be used to store ZODB data in other media or formats, as we'll see later in more detail.
Zope Enterprise Objects (ZEO) is a network storage for the ZODB. Using ZEO, any number of ZODB clients can connect to the same ZODB. ZEO can be used to provide scalability because the load can be distributed between several ZEO clients instead of only one.
ZODB and relational databases
By far, the most popular mechanism for storing program data is a relational database. The relational model uses tables, which contain data that conforms to a predefined schema. Each table column represents an attribute of the schema and each row is a set of values for those columns. The power of this model comes from the ability to relate tables by one or more common attributes, so that data can be queried and assembled in multiple ways.
Relational databases are widely popular, in part because their programming language independece makes them relatively easy to use in a variety of work environments. They usually require specific drivers for any given programming language, but that's not really a problem in the case of Python, as it has bindings to all the major relational databases (and some minor ones).
Of course, the vast majority of relational databases do not natively store Python objects, so it's necessary for the application itself to read the data from the tables and "assemble" the columns from each row into the required objects. The application is also responsible for breaking apart the objects and fitting their attributes into the table structure when a change is detected.
These assembly and disassembly is known as object/relational mapping and can use a significant portion of an application's logic. Fortunately, there are excellent third party Python packages, called ORMs (object-relational mappers), that take care of the interaction with the relational database for the application developer.
Using an ORM allows the developer to forget about the underlying database and focus on the Python objects, but the objects themselves most retain the tabular structure. Also, in many cases it's necessary to have a very good understanding of the specific database used and relational databases in general to be able to decide how best to structure the objects.
Working with the ZODB does not require any of this mapping activity, since the objects are stored in their native form, which simplifies the interaction between the developer and the data. Without the need for a tabular structure, data can better reflect the organization of information in the problem domain.
Instead of managing relations using different tables with common primary keys, the ZODB lets developers use normal Python object references. An object can be a "property" of a separate object without the need for table joins and multiple objects can reference this property without each actually having to store a copy of the object.
Because it works directly with Python objects, the ZODB doesn't require a pre-defined structure of columns and data types for the objects it stores, which means that object attributes can easily change both in quantity and type. This can often be a lot harder when using a relational database for storage.
One other advantage of using the ZODB over a relational database comes when the problem domain requires a filesystem-like structure. Modeling this kind of containment relationships does not come naturally for the relational model, but is quite easy with the hierarchichal nature of the ZODB. Content management systems are one example of an application domain that is very well suited for ZODB use.
Is the ZODB a NoSQL database?
In recent years, the term NoSQL has been consitently used to refer to a "new" breed of database systems which basically do not use the relational paradigm. Here is one semi-official definition of NoSQL, taken from http://nosql-database.org/:
"Next Generation Databases mostly addressing some of the points: being non- relational, distributed, open-source and horizontallly scalable. The original intention has been modern web-scale databases. The movement began early 2009 and is growing rapidly. Often more characteristics apply as: schema-free, easy replication support, simple API, eventually consistent / BASE (Basically Available, Soft state, Eventual consistency, or not ACID), and more."
The ZODB has been around for more than a decade and thus clearly predates this concept (as do most of the NoSQL databases in existence), but in the general sense it can be classified as a NoSQL database, because it shares the main characteristic of being non-relational.
The ZODB is also open source, horizontally scalable and schema-free, like many of its NoSQL counterparts. It is not distributed and does not offer easy replication, at least not for free.
ZODB != Zope
Zope is a web application server written in Python that has also been around for more than 10 years. Unlike most web frameworks, Zope encourages the use of an object database for persistence, rather than the usual relational database. The database used by Zope is, of course, the ZODB. The ZODB has been a vital part of Zope since Zope's creation, as you may have already guessed by its name.
In part for its strong association with Zope and probably also in part due to the low popularity of object databases in general, the ZODB is used very little outside of the Zope world. Developers without exposure to Zope tend to assume that you have to use one to get the other or are afraid that they would have to pull dozens of Zope dependencies if they chose to use the ZODB. Some might even believe that they have to write code in the 'Zope way' if they want to use it.
Part of the motivation for writing this book is to clearly show the wider Python world that the ZODB is a totally independent Python package that can be a much better fit than relational databases for data persistence in many Python projects. The ZODB is sufficiently transparent in use that you only need to follow a few very simple rules to get your application to store your objects. Everything else is "just Python".