Graduated MSc Computing Science
“Change Impact Analysis for Rebel Specifications”
Intro - In software development, ideas and abstractions meet to evolve into applications. For large companies this is often a process that does not allow for mistakes that result in a loss of any kind. To be assured that applications are unambigu- ously aligned with their intentions, formal methods can be used. Formal (specification) languages are part of these meth- ods and make it possible to write specifications (of entities) for which its properties can be proven correct. After con- struction, specifications may be utilized for automated code generation, documentation and other software engineering purposes. …
This thesis finished my MSc Computing Science at Utrecht University. I wrote it together with ING, which uses Rebel to model its financial products.
Rebel and the problem with changes
Rebel is a formal specification language that ING uses to describe concepts from banking, like accounts and transactions, in a way that leaves no room for interpretation. Application code can be derived from such a specification in a provably correct way. Requirements change over time though, and so do the specifications. A change can alter the set of states a specification can be in, its state space, and an application derived from the new version could then behave in a way nobody expected, even though it was derived correctly. For a bank, where downtime or lost data is not an option, you want to know the impact of a change before it goes anywhere.
The question of the thesis is how formal methods can be used to compute that impact. There was no complete description of Rebel’s syntax and semantics yet, so I first wrote both down, with the semantics based on extended finite state machines. After comparing several formal specification languages I picked the B-Method and built a translation from Rebel to it.
Computing the impact
A change is written as an invariant, a property that has to hold in every state of the model, so you can express it without knowing the whole model. Two methods update the model with that invariant: a direct one, which is quick, and an iterative one, which tells you precisely where the invariant is violated. The model checker ProB then explores every reachable state of the model before and after the update, and the difference between those two state spaces is the impact. Bash scripts glue the steps together around ProB’s command line tool.
The experiments showed that these state space differences give a usable indication of the impact of a change before it is applied, where there was no information about it at all before. They also showed where the approach stops: ProB checks every state explicitly, the number of states explodes as models grow, and larger models could not be explored in a reasonable amount of time. Part of the thesis compares ProB’s options, symmetry reduction among them, to keep that state space as small as possible.
|