Food for thought : http://spinroot.com/p10/ (2006)

From people who are actually writing code for spaceship and exploration robot.

However, I am not sure how applicable are these kind of advices to other domains. I am working on critical medical devices, so we are already following most of these recommendations, but if you are designing a website managing millions of users, the advices would look totally different. It would be silly to write everything from scratch in C without any kind of dynamic memory allocation.

I am not convinced the former is necessary better or more challenging than the latter, it is just a different context.

The criteria for software quality that this is mainly about are reliability and efficiency (https://en.wikipedia.org/wiki/Software_quality#Measurement). I think these are important for all software domains.

Margaret Hamilton [0] coined the term "software engineering" to describe the standard to which for Apollo had to be designed, written and tested. It was a high standard because her team was embedded in a larger engineering culture of engineering excellence where shortcuts and being in love with one's own ideas meant people might die needlessly.

Adopting that attitude may be a place to start.

Good luck.

[0]: https://en.wikipedia.org/wiki/Margaret_Hamilton_(scientist)

Around 26 years ago, I developed firmware for geosynchronous satellites in a "safe" subset of Ada87. I still think it's a wonderful language and wish there were more opportunities to develop with it. Since then, I've worked solely in C and assembly.

Money. Time helps.

Creating reliable software-intensive systems is something we know how to do, we have repeatedly done it in the past, and we are doing in the present. Only problems are it ain't cheap in terms of money or time, so to be cost effective, the costs associated with failure or malfunction must be high.

Safety critical systems (where human health can be endangered by system failures and malfunctions) are one such system. They're usually developed under some sort of regulatory oversight (e.g., FAA, FDA). There are a set of minimal government standards that must be met, much of these related to documentation, and much of the documentation relating to risk analysis and quality management. There are also standards administered by industry groups---often, these are "voluntary" in name but for liability reasons practically required (in my experience). There's going to be some sort of (mandated) monitoring of the product after release (which feeds into the quality system) to correct design flaws, defects, etc. You have regulatory affairs people on staff to handle this area. The safety critical system itself is developed under some sort of formal product lifecycle management system. There will be a formal set of requirements. Early in the process, there will be a risk analysis (continually updated) of the design. There will be an a priori development plan, timeline, and cost estimates. As for development models, the waterfall model is alive and well, in part because it coöperates with the demanding documentation requirements; waterfall does seem to be on the decline in favor of more agile models (but not too agile!).

The software itself may be assembler if you're really resource constrained, but more likely, it will be C or C++ with a restrictive coding standard, or real-time Java, or possibly Ada. If you're fancy (and progressive), there might be a system model with automatic code generation. The tooling you adopt (e.g., from Atego) will assist with verification and validation, as well as requirements traceability. Most people, I think, find software work under these conditions to be boring and frustrating---it's certainly limiting. You're maybe generating 100 SLOC a day? The system is almost certainly running on some sort of real-time OS like vxWorks or QNX, and the hardware, software, and any associated libraries are (as much as possible) validated for use for your application. You can, given the right circumstances, use COTS software (a.k.a. SOUP, software of unknown pedigree), but then the validation becomes your responsibility.

The HN crowd hates it, but in my experience on these sorts of projects, degrees and certifications are required and expected except at perhaps the very lowest level of employees. Electrical and mechanical designs are probably stamped by a real engineer; depending on the nature of the software portion (and where it's being developed) the software design might get stamped too, but more likely, it's being supervised by someone with a bunch of industry-relevant certifications.

I've glossed over a ton, and skipped even more.

In my opinion: I think that the most important lesson to learn from safety critical projects (and good engineering practice in general) is that tools are not the major driver of quality: proper engineering processes and quality management are the driver of quality. A "right-sized" product lifecycle management plan, including a "right-sized" quality system are critical and necessary to predictably and reliably produce a quality product. All else being equal, I'd predict that a C application developed under a lightweight PLM and QMS will be superior to one developed under Agile and Rust, or Scrum and Haskell, or whatever the du jour fad might be.

I agree on 90% of your post and detailed description.

However, on the conclusion, I am fairly convinced that the process is necessary but not enough to produce reliable system. I would say the most important parameter is "People".

If you have a team where people trust each other, where you can report odd issues without fear of reprisal, where people really want to deliver something safe, it will work, otherwise, with a dysfunctional team, using the exact same process, you'll get a terrible mess.

I did neglect people and am wrong in doing so; I think what you say is absolutely correct.