This guide explains how to write Python programs that use the Z Object Database (ZODB) and Zope Enterprise Objects (ZEO). The latest version of the guide is always available at

What is the ZODB?

The ZODB is a persistence system for Python objects. Persistent programming languages provide facilities that automatically write objects to disk and read them in again when they’re required by a running program. By installing the ZODB, you add such facilities to Python.

It’s certainly possible to build your own system for making Python objects persistent. The usual starting points are the pickle module, for converting objects into a string representation, and various database modules, such as the gdbm or bsddb modules, that provide ways to write strings to disk and read them back. It’s straightforward to combine the pickle module and a database module to store and retrieve objects, and in fact the shelve module, included in Python’s standard library, does this.

The downside is that the programmer has to explicitly manage objects, reading an object when it’s needed and writing it out to disk when the object is no longer required. The ZODB manages objects for you, keeping them in a cache, writing them out to disk when they are modified, and dropping them from the cache if they haven’t been used in a while.

OODBs vs. Relational DBs

Another way to look at it is that the ZODB is a Python-specific object-oriented database (OODB). Commercial object databases for C++ or Java often require that you jump through some hoops, such as using a special preprocessor or avoiding certain data types. As we’ll see, the ZODB has some hoops of its own to jump through, but in comparison the naturalness of the ZODB is astonishing.

Relational databases (RDBs) are far more common than OODBs. Relational databases store information in tables; a table consists of any number of rows, each row containing several columns of information. (Rows are more formally called relations, which is where the term “relational database” originates.)

Let’s look at a concrete example. The example comes from my day job working for the MEMS Exchange, in a greatly simplified version. The job is to track process runs, which are lists of manufacturing steps to be performed in a semiconductor fab. A run is owned by a particular user, and has a name and assigned ID number. Runs consist of a number of operations; an operation is a single step to be performed, such as depositing something on a wafer or etching something off it.

Operations may have parameters, which are additional information required to perform an operation. For example, if you’re depositing something on a wafer, you need to know two things: 1) what you’re depositing, and 2) how much should be deposited. You might deposit 100 microns of silicon oxide, or 1 micron of copper.

Mapping these structures to a relational database is straightforward:

  int      run_id,
  varchar  owner,
  varchar  title,
  int      acct_num,
  primary key(run_id)

CREATE TABLE operations (
  int      run_id,
  int      step_num,
  varchar  process_id,
  PRIMARY KEY(run_id, step_num),
  FOREIGN KEY(run_id) REFERENCES runs(run_id),

CREATE TABLE parameters (
  int      run_id,
  int      step_num,
  varchar  param_name,
  varchar  param_value,
  PRIMARY KEY(run_id, step_num, param_name)
  FOREIGN KEY(run_id, step_num)
     REFERENCES operations(run_id, step_num),

In Python, you would write three classes named Run, Operation, and Parameter. I won’t present code for defining these classes, since that code is uninteresting at this point. Each class would contain a single method to begin with, an __init__() method that assigns default values, such as 0 or None, to each attribute of the class.

It’s not difficult to write Python code that will create a Run instance and populate it with the data from the relational tables; with a little more effort, you can build a straightforward tool, usually called an object- relational mapper, to do this automatically. (See for a quick hack at a Python object-relational mapper, and for Joel Shprentz’s more successful implementation of the same idea; Unlike mine, Shprentz’s system has been used for actual work.)

However, it is difficult to make an object-relational mapper reasonably quick; a simple-minded implementation like mine is quite slow because it has to do several queries to access all of an object’s data. Higher performance object- relational mappers cache objects to improve performance, only performing SQL queries when they actually need to.

That helps if you want to access run number 123 all of a sudden. But what if you want to find all runs where a step has a parameter named ‘thickness’ with a value of 2.0? In the relational version, you have two unappealing choices:

  1. Write a specialized SQL query for this case: SELECT run_id FROM operations WHERE param_name = 'thickness' AND param_value = 2.0

    If such queries are common, you can end up with lots of specialized queries. When the database tables get rearranged, all these queries will need to be modified.

  2. An object-relational mapper doesn’t help much. Scanning through the runs means that the the mapper will perform the required SQL queries to read run #1, and then a simple Python loop can check whether any of its steps have the parameter you’re looking for. Repeat for run #2, 3, and so forth. This does a vast number of SQL queries, and therefore is incredibly slow.

An object database such as ZODB simply stores internal pointers from object to object, so reading in a single object is much faster than doing a bunch of SQL queries and assembling the results. Scanning all runs, therefore, is still inefficient, but not grossly inefficient.

What is ZEO?

The ZODB comes with a few different classes that implement the Storage interface. Such classes handle the job of writing out Python objects to a physical storage medium, which can be a disk file (the FileStorage class), a BerkeleyDB file (BDBFullStorage), a relational database (DCOracleStorage), or some other medium. ZEO adds ClientStorage, a new Storage that doesn’t write to physical media but just forwards all requests across a network to a server. The server, which is running an instance of the StorageServer class, simply acts as a front-end for some physical Storage class. It’s a fairly simple idea, but as we’ll see later on in this document, it opens up many possibilities.

About this guide

The primary author of this guide works on a project which uses the ZODB and ZEO as its primary storage technology. We use the ZODB to store process runs and operations, a catalog of available processes, user information, accounting information, and other data. Part of the goal of writing this document is to make our experience more widely available. A few times we’ve spent hours or even days trying to figure out a problem, and this guide is an attempt to gather up the knowledge we’ve gained so that others don’t have to make the same mistakes we did while learning.

The author’s ZODB project is described in a paper available here,

This document will always be a work in progress. If you wish to suggest clarifications or additional topics, please send your comments to the ZODB-dev mailing list.


Andrew Kuchling wrote the original version of this guide, which provided some of the first ZODB documentation for Python programmers. His initial version has been updated over time by Jeremy Hylton and Tim Peters.

I’d like to thank the people who’ve pointed out inaccuracies and bugs, offered suggestions on the text, or proposed new topics that should be covered: Jeff Bauer, Willem Broekema, Thomas Guettler, Chris McDonough, George Runyan.