3 ms·
If the scope is very narrow, software can be engineered to be correct. However, I think it is better to work towards organic, self-healing software systems like
by wallflower 17y ago
If the scope is very narrow, software can be engineered to be correct. However, I think it is better to work towards organic, self-healing software systems like Google (where they leave failed cluster nodes in place, not bothering to find or remove them).
Another reason software cannot be engineered in a cost-effective manner is most engineering is based on inviolable physical rules (the constant g (gravity) - this allows software (irony, yes) to be developed that helps engineers correctly (there is always a "factor of safety") engineer (FEA software like ANSYS). My hat is off to the legends like Thomas Roebling who designed the Brooklyn Bridge before computer aid.
The pain and cost of engineering software to structural engineering standards (ISO 9000 is a joke, NASA has its own internal standards) usually exceeds the lifetime benefit of correctly engineering that software. A good rule of thumb is will someone's life be jeopardized if the software fails (dialysis, automobile vehicle systems, nuclear thermal rod control, 747s, space ships (makes me wonder, actually, how "correct" Burt Rutan's X Prize spaceship software may be)).
- hxa7241 17y ago> most engineering is based on inviolable physical rules Don't rules such as those of algorithmic complexity have the same status and position? They are as objectively certain -- more so, in fact, since they are purely logical. At its base, software is built on bits and operations, which are entirely determinate. The ramifications might not currently be fully understood or exploited (software is only about 50 years old!), but the potential is there. As an example: with CPUs like the Z80 it was possible to count clock cycles of instructions, and determine upper/lower/average running times of code. That would be equivalent to determining resultant forces across a whole physical structure. Today, CPU makers have made things difficult, but that is completely contingent: it would in principle be possible to predict performance, in an engineering fashion, for many purposes.
- wallflower 17y agoIn the homework we do for engineering classes, we would be able to make simplifications without fundamentally altering the problem. For example a continuous load on a structure could be modeled as a single force vector because G, the gravitational force always acts downwards. You can't do that in software with loosely-coupled, autonomous modules. These modules do not obey anything. Just an implicit contract that can be broken easily. Unit tests only help to clarify the contract for a module or between modules. In structural engineering, there are a lot of unknowns but they can be modeled across the entire range because there are many known constants. The load-bearing capacity of the beam, the moment (the twisting force acting on the beam). Virtual abstractions like software do not obey physical rules. Because they are virtual. You can create a Castle in the Air in your program world, if you like. Hard physical reality has real constraints. Virtual reality has flexible constraints. Physical reality is one of the reasons building construction is a solved problem. The irony is with sophisticated software, architects like S. Calatrava and F. Gehry can begin to create almost Castles in the Air.