[EuroPython 2012] Arnaud Fontaine - 3 JULY 2012 in "Track Lasagne"
Recent cloud architecture are heavily based on key-value storage databases rather than distributed filesystems. While the data itself is usually stored in scalable database, the source code is mainly kept on the filesystem. However, having the source code on filesystem raises a lot of issues for cloud applications and more generally for transactional web applications supposed to be deployed on a large number of servers: Everytime the source code is modified, it has to be synchronized on all the nodes somehow and the application server must be restarted, therefore implying downtime or inconsistency. Usually, the developer cannot directly modify the source code as this task is left to the administrator. This talk will show how source code, including modules and their classes, can be migrated away from filesystem, thanks to import hooks, meta classes, reflection and metaprogramming provided by Python, and the different challenges involved: Technical: modules and classes must be loaded and reset while carefully considering concurrency and database issues and ensuring that the system will stay usable whatever happens and without breaking existing class instances. Also, developers must be efficient with the new system, including convenient source code edition and meaningful backtraces. Performance: having dynamic modules and classes involved a performance tradeoff which should be kept to the minimum. Security: as the source code is dynamically loaded from the database, extra care must be taken care of compared to filesystem source code where the source code is never supposed to change once the applications has been started. In order to give practical examples of how this migration can be achieved, two examples will be studied and explain thoroughly, based on the experience I have acquired through the implementation for ERP5 I have worked on during the past year: First, through a basic WSGI application based on a lightweight web application framework to explain the basics behind having Python source code available in a transactional key-value storage database and more precisely how modules and their classes can be dynamically loaded and reset on-demand upon modifications while keeping good performances. This usually involved specific features from the database point of view which will also be discussed. Secondly, through ERP5 which is mainly known as an open source ERP based on Zope and thus ZODB, and as such it provides features an ERP typically provides, but it is also used for a wide range of other purposes. Therefore, this will give a more real-world example on how this migration has been performed on a complex web application where 90% of the code is going to be moved away from filesystem into the ZODB.