The state on a later Level C & D project was in worse shape: bloated architecture, lots of requirements churn, little to no peer review, one unrealistic deadline after another, and mandatory overtime for large periods. It was not an environment conducive to thoughtfulness or quality. I finally got fed up and quit.
The standard does mention "formal verification/methods/proof" but to my knowledge it's rarely been used extensively.
The actual failure rate of software to this standard seems to be at least two orders of magnitude higher.
What made it better:
Concurrent editing. Only one person at a time was modifying any section, but two or more sections could be edited by a group of people.
Version control baked in. VC is awesome, reducing friction means that it actually gets used. Checking in/out sections was just part of the process.
Automatic traceability matrices. These are not easy to make by hand. That way is error prone and tedious. It's also hard to verify. With DOORS we were able to generate these automatically. Verifying them was relatively painless because we weren't flipping through several 1k page documents to make sure things actually matched, we could easily skip to only the applicable parts of any document. The main error that it didn't reduce was when a requirement didn't get linked to every test case or design feature that it should've been linked to.
Again, probably better tools out there than DOORS (at the time, or hopefully by now). But it's a lot better than what most offices do with either post hoc document generation or ad hoc generation with Word + Excel.
Sure sounds like the way it should be to me. Maybe it needs to be made easier to prove your software is correct, but to me it seems like for systems like airplanes, code that cannot be proven correct is worthless.
Considering most programmers are said to produce 6 lines of good code a day, maybe it's not even actually slower in the end if the formal verification process filters out every other line you would have written that day.
the apologetic attitude needs to stop. programs are just complicated machines and should be treated as such not some mystical voodo that will be broken no matter what
However it's a complete different matter for civil airplanes : it's the lead dev/project manager which is liable for life for what he has shipped. For example, a retired engineer from Airbus was heard in trial for the Concorde accident in France in 2000.
Formal verification and state space analysis can prove that the software "model" will not fail. State space exploration of the actual implementation is actually often not feasible due to the enourmous amount of states.
So my question: Are you doing formal analysis of the software models/designs? I know they are using it in software used in space crafts / nuclear power plants.
Reference: - http://ti.arc.nasa.gov/m/profile/dimitra/euromicro-share.pdf - http://javapathfinder.sourceforge.net/
For example Astree [1] has been developed for decades now, with Airbus as one of the major sponsors.
On the other hand, I also wonder why people don't think of using Erlang for something like this. The VM is designed for a ridiculous amount of uptime, it has a supervisor tree that can monitor and restart failed processes, and it can interface with C/C++. A rock-solid VM should be running and monitoring life-safety systems.
I'll speak to the 787 avionics system. It used a system of channels/buffers and processes very much like what Erlang and Go use for interprocess communication (I'm trying to remember now if channels could be received in multiple processes like Go or if only a single process could receive like in Erlang). This was an excellent model for what we were doing, and really for a lot of systems this sort of CSP and actor style model maps well.
What alternative are you suggesting?
AFAIK (partial) assurance in C/C++ can only be handled by additional testing tools, Frama-C for instance.
I agree that C/C++ should not be used for security applications. Ada is a much better choice because it was designed for security.
In fact, they (http://windriver.com/products/product-overviews/PO_Diab_Comp...) say:
Diab Compiler has been a reliable code generation tool for
avionics products certified for DO-178B, products for the
nuclear market certified to IEC 60880, railway applications
certified to EN 50128, and industrial products certified
to IEC 61508, and is now qualified for use in automotive
applications certified to ISO 26262.
Ada does have some built-in advantages, but I think my point still stands: the language is a small part of the entire SDLC, and I don't think it's the most important part.I'm pretty sure that the only industrial formally verified compiler is CompCert (for C), though I could be wrong. The motivation for CompCert was certainly that Airbus wanted such a compiler.
Ada wouldn't be a better choice simply because it's designed for security. It'd be a better choice if it turned out better in practice. I've read some of the studies that have been done, and I haven't found them convincing.
Requiring additional tools just isn't a problem, if it works well. Don't criticise the process, criticise the result.
The focus on languages instead of the SDLC is telling, I think.
I once worked on satellite systems using Ada95 and Ada2005 (Ada is definitely not dead). The language is a pain to use but is impressive in that it catches more crap at compile-time than anything else I've seen.
True. More information: http://www.ada2012.org
For a quick example, there are many languages where you cannot accidentally run off the end/start of an array, barring a compiler error.
With a language like C / C++, it's possible. Not probable, given that sort of testing. But possible nonetheless.
Some languages are also easier to test than others, partially because of this, partially because of other issues. For instance, in some languages you can guarantee at the language level (again, barring compiler errors) that something won't be modified. (Like const, but actually working.)
I wonder if someone can pull an #ElonMusk on the DO-178 to slim things down in order to have better control.
ps: planes fly with bugs, see the DreamLiner, Airbus ones aren't free from them either, employees know this. (I guess they travel by train)
The train brake control systems are written in ... C.
All safety critical software (every piece of code ran on-board is safety critical the least) in aerospace needs to pass the DO-178 standard [1].
That is far more serious than standard unit tests you are used to in node.js applications. Generally speaking, to develop a piece of code under that standard it takes 20% of time to write the code, and 80% to testing, and enormous amount of documentation (that is optimistic estimation, usually worse).
Quoting speaker from DO-178 training course I attended:
If this crash confirms the cause to be a software bug, that is something much bigger than an airplane crash - a huge punch to the whole federal aviation administration.[1] - http://en.wikipedia.org/wiki/DO-178B