Python Success Stories
Nexedi is a leader in high-end enterprise services, providing solutions for Enterprise Resource Planning (ERP), Customer Relationship Management (CRM), and eCommerce. Nexedi has built its business on Open Source, and has designed and released an ERP/CRM framework called ERP5 under the GPL Free Software license.
ERP5 is in production in the apparel industry and government agencies with multi-gigabyte databases that track millions of warehouse stock movements. ERP5 is written entirely in Python and leverages the Zope Enterprise Objects framework to provide high performance and availability on clusters of inexpensive PCs.
The ERP5 project started in 2002 when Coramy, an apparel industry leader in Europe, decided to migrate its in-house ERP solution to a new open source model. The choice of open source for Coramy ERP was a strategic move to reduce software maintenance costs and to allow Coramy to retain complete control over its custom developments, something that would have been impossible with standard proprietary ERP solutions.
Nexedi was created as an independent company in charge of developing, implementing, and disseminating the ERP5 technology. Nexedi was given a budget of 80,000 EUR to develop a generic ERP framework published under GPL license and customize it for Coramy's specific needs.
Python: Clean Objects from Scripting to Metaprogramming
Working with this small budget, ERP5's developers needed to look to innovative approaches and code reuse to cut development costs. At the most abstract level, ERP5 is based on five generic classes used by all modules: Resource, Node, Movement, Item, and Path. This model, known as the ERP5 Universal Business Model, makes it possible to reuse code by abstracting away from the specific domain and encapsulating the generic relationships and actions common to many business processes. As a result, modules as different as Payroll and Invoice can share almost of all their code.
In this example of the ERP5 Universal Business Model, Movement allows the transfer of a quantity of resource from one node to another, Path allows planning of resource sourcing, and Item provides Resource traceability. Zoom in
The ERP5 abstract model architecture requires a well-designed object language that supports high level abstraction. Selecting an appropriate implementation language became strategic to the project's success. The project also required complete multi-platform OS abstraction, a rich library for Web applications, a rapid GUI development toolkit, support for internationalization, wide community support, and proven maturity. The short list of candidates quickly became: Java or Python. Or, more precisely, Java+Jakarta or Python+Zope.
Two key factors led to the choice of Python. First, was the need to make extensive use of metaprogramming. Second, for simplicity's sake, ERP5 needed to be implemented in a single language from core architecture to scripting.
Metaprogramming is a technique that allows the programmer to redefine the semantics of the implementation language at runtime. It can be used to endow extremely abstract implementations with domain-specific behaviors that are modeled in properties or tables, rather than hand-coded. In ERP5, this powerful technique allows 95% of the class methods to be generated automatically from lists of properties that define each unique custom ERP implementation. This has reduced maintenance costs by an order of magnitude: The typical ERP5 system contains 100,000 lines of code, rather than the 1,000,000 lines of code required in similar projects based on traditional programming techniques.
Python supports efficient metaprogramming through powerful introspection features that allow programs to inspect and modify code at runtime. Java's introspection are by contrast quite poor and inflexible. Metaprogramming in Java requires writing preprocessors, which would have added prohibitively to the cost and complexity of implementing the ERP5 system.
Another advantage that Python had over Java was that it could be used at all levels of the system, from core implementation to scripting. Most ERP systems, while written in one language, use another scripting language to allow flexible configuration at run time by ERP administrators. Python is equally well suited both for scripting and core development, reducing complexity and increasing the flexibility of the system. Using Python allowed code initially written as scripts to be incorporated afterward into core components, and vice versa, wherever this made sense. With Java, it would have been necessary to provide a separate scripting environment based on a different language, such as Jython or ECMAScript, and reusing scripting code in core components would have been much more difficult.
Enterprise Objects with Zope
In addition to leveraging these high-level language capabilities, the ERP5 project also needed to base on an existing turn-key application server with support for transparent object persistence, transactions, and workflow. This would allow the project to focus its limited resources on application design, rather than on application server design.
Zope, a Python-based application server, fulfilled this need. In 2002, Zope was already a mature application server environment while Jakarta was still rather immature. Zope provided clustering, object storage, object publication, transactions, security, workflows, and a web-based management interface, all in a turn-key package. Zope and Python, unlike Java, also provided the licensing needed for complete freedom in distributing code on LiveCD, in RPM packages, and so forth.
In some applications needed in ERP, such as in Point-of-Sales (POS), an autonomous client/server GUI application provides better results than a pure Web-based solution. For these cases, the ERP5 project team selected PyQt, which supports the rapid development of multi-platform client/server applications with native look and feel. Autonomous applications written in PyQt could interoperate with Zope through the XML-RPC and SOAP implementations available for Python.
ZSQLCatalog: Querying and ROLAP of Zope Object Database
In January 2003, the initial ERP5 modules went into production at the Coramy factory and design center.
The ERP5 architecture relies completely on Zope for data storage, transactions and workflow management. The user interface is based on an extended version of Martijn Faassen's Formulator component. ERP5 itself is implemented as a set of Zope components. Overall, thanks to massive code reuse and fast coding, ERP5's initial development was completed in less than one year. The choice of Python and Zope proved to be a good one.
Part of this success was due to Nexedi's ZSQLCatalog component which leverages Zope's object database to implement an innovative approach for querying objects and data mining.
ZSQLCatalog solves several major problems often encountered in Enterprise applications: locks, slow reports, and inconsistent transactions. Most Enterprise applications use tables in a relational database to store data. In ERP5 data is stored instead in Zope's object database. Zope eliminates the need for storage adapters or attribute mappings by providing transparent persistence of Python objects. Zope also speeds up object access: Reading the Zope object database is 10 to 100 times faster than retrieving a row from the fastest relational databases available on the market today. Finally, unlike relational databases, transactions in a Zope object database can last minutes without causing deadlocks.
One limitation encountered in Zope was that it does not provide an efficient tool for querying the millions of objects that are needed in an ERP solution. To solve this, Nexedi's ZSQLCatalog maps between attributes or methods of objects and relational columns, tables, and databases. This mapping is not intended to store actual object attributes into relational tables, but rather only those attributes relevant to facilitating fast queries. In this way, the relational database is used as an index into the Zope object database.
To provide support for fast queries of the object database, ZSQLCatalog copies select attributes (dots) and relations (lines) from the object database into tables of a relational database. Zoom in
ZSQLCatalog can be viewed as a kind of Relational Online Analytical Processing (ROLAP) component for object databases, and it provides a nice interface for extracting reports from the object database into OpenOffice or Microsoft Excel spreadsheets.
ZSQLCatalog was very successful in this project: a Zope database with more than 2,000,000 objects can be queried with statistical methods in a few milliseconds.
ERP5SyncML: Synchronized Distributed Objects
Coramy's requirements for ERP5 included the need to deploy applications in its factories in Tunisia. This posed some additional challenges, since internet connectivity between Africa and Europe is not always perfect and can be very expensive. Given Coramy's modest budget for network connectivity, the project had to provide a solution for synchronizing two ERP5 servers over slow and unreliable transcontinental internet connectivity.
To do this, Nexedi implemented the SyncML protocol in Zope, using email or http as the transport layer. Early prototypes were developed in a few weeks, using Python's rich library for network protocols and XML parsing.
Turning those prototypes into an industrial-strength solution turned out to be a bit more complex, due to the many possible failure cases the code had to handle. Nexedi used Python's unit testing framework to write extensive tests of the SyncML implementation. These tests could test every part of the SyncML component, especially those that would be difficult to test in the field, and report any problems back to the developer. Once this was done, the ERP5SyncML component became reliable. In the process of testing, some minor bugs occurring in complex situations were found in the Python libraries. Because Python is open source, the ERP5 developers could find and fix the bugs quickly and contribute the fixes back to the Python community.
ERP5SyncML is now being used not only by ERP5 but was also adopted by the Nuxeo CPS, a Content Management System, as a way to synchronize documents between French government agencies.
Python-GLPK: Linear Programming in Python
Some aspects of an ERP system's database query functionality require mathematics that go beyond the capabilities of the relational query model. For example, ERP5 uses linear programming to determine resource capacities. Although Python includes excellent numerical frameworks, C or FORTRAN implementation of complex scientific algorithms is usually much more efficient.
Nexedi found the GNU Linear Programming Kit (GLPK) to be a good starting point for ERP5's linear programming needs. GPLK is written in C, and interfacing it to Python was achieved in only a couple of hours using the SWIG glue libraries. Nexedi now distributes a Python GLPK module, python-glpk, which provides the power of linear programming in Python.
The ERP5 abstract model has been found to reduce development costs by an order of magnitude when compared to traditional ERP architectures, and it has performed well in large mission-critical Enterprise applications. ERP5's largest system runs on a cluster of eight CPUs. It serves fifty simultaneous users, each of them with eight parallel sessions, and it handles more than 2,000,000 Python objects. Its ZSQLCatalog relational index holds more than 10,000,000 rows.
Python and Zope were key to this success. Python provided a powerful object language and a rich set of libraries which allowed quick development of clean and compact code. Zope provided a mature application server and object database.
Nexedi is sometimes asked: Why not Java and J2EE? While it would be possible to create a similar system with Java and J2EE, development costs would be much higher. Jakarta and ObjectWeb have both matured but are still poorly integrated when compared with Zope, and require a much more complex development style. Java's poor introspection features are also still a serious limitation for efficient metaprogramming, for using Java itself as a scripting language, or for flexible object persistence. Using Java is simply not consistent with today's trend of cost cutting in Enterprise development. If the choice were made again today, Nexedi would still opt for Python and Zope.
In June 2004, ERP5 was nominated for "best enterprise project" by Decision Informatique professional magazine. Nexedi is now working to simplify the ERP5 setup and configuration process, in order to ease its adoption by a wider audience of open source developers.