5 ms·
I've worked on software for some time and have seen a wide range of interesting failure modes. All of the interesting failures (some of which have had large mo
by dekhn 7y ago
I've worked on software for some time and have seen a wide range of interesting failure modes. All of the interesting failures (some of which have had large monetary or visible outage impact) involved systems that broke in ways that aren't well-addressed by proof systems. Instead, they've been about things like statistical binning of parts, rare errors that escaped through middling checksums, and complex distributed systems. In most of these cases, there is some physical component, such as corrupted packets which pass CRC checksum, or a chip that occasionally miscomputes the hash of a string. Not sure how proof systems handle this, but in practice, those have been some of the most impactful errors, and they are typically dealt with using heavy but not comprehensive software testing.
- lmm 7y agoIn my experience the interesting failures are a tiny fraction of the failures; your mind just glosses over the vast sea of boring failures. The most impactful software flaws have not been subtle logic misalignments, they've been utterly basic errors like Cloudbleed or the Arianne 5 overflow. Maybe proof systems can only address 99% of software failures; my guess is that with experience we'll find ways to cover the rarer cases as well, but even if we don't, they're still worth using.
- jwhitlark 7y agoWell, maybe. Additional costs with these approaches are often ignored. I wouldn't want to give up proof systems as a tool, but too many treat them as a zero cost panacea. A proof system or type checker is not an excuse to not give a fk about your craft, but I've seen it that way more frequently than I'd like to admit. Personally and honestly, outside of already well understood or defined systems, I'd say that people should invest their time in understanding their business domain, before they dig deep into proofs and related techniques.
- lmm 7y ago> Well, maybe. Additional costs with these approaches are often ignored. I wouldn't want to give up proof systems as a tool, but too many treat them as a zero cost panacea. A proof system or type checker is not an excuse to not give a fk about your craft, but I've seen it that way more frequently than I'd like to admit. The acceptable defect rate to the business is presumably fixed, so if a proof system or type system reduces you defect rate below that then that should be an "excuse" to write fewer tests and move faster, as icky as it feels to do that. I find the maintenance burden of tests is often underestimated; types in particular may not be free, but they get you more bang for your buck than any alternative, in my experience. > Personally and honestly, outside of already well understood or defined systems, I'd say that people should invest their time in understanding their business domain, before they dig deep into proofs and related techniques. Isn't it the same thing though? Encoding your business domain into the system you're using is the hardest part of using any formal system, but it's also an excellent way of forcing you to really understand the domain.
- jwhitlark 7y agoIn my experience, people either truely care about the business domain, and are willing to be uncomfortable to delve into it, or not. Types and proofs seem to be orthogonal. Some domains are worth it, some are not. Sometimes it fits well, sometimes it doesn't. Just don't be the one sinking into the muck saying, "...but the math worked". The math will be far more reliable than your understanding of the domain.
- dekhn 7y agoI agree, the vast majority of failures I see are "operation timed out", for utterly banal reasons. But I'd like to point out that the "impact" of Ariane 5 was very different from the "impact" of Cloudbleed. Space operations are definitely a situation where large amounts of advance planning and testing makes a huge difference, because you only get one shot to put a $500M satellite in orbit.
- Palomides 7y ago>complex distributed systems amazon has been using TLA+ to prove/check correctness of production distributed algorithms and systems for about a decade
- dekhn 7y agoSure, here's an example: https://lamport.azurewebsites.net/tla/formal-methods-amazon.pdf https://lamport.azurewebsites.net/tla/formal-methods-amazon.... They still have to let users know about rare bugs that led to data corruption. You'll generally formal proof checking applied to simplified model systems of individual components, while interesting failures tend to involve ten or more components.
- redis_mlc 7y ago> amazon has been using TLA+ to prove/check correctness of production distributed algorithms and systems for about a decade Such an in-your-face, smug and flippant comment but ... On some projects. In some teams. And Aurora lost a lot of data initially in spite of whatever TLA+ they were using, somewhere at Amazon.