5 ms·

I think I see what you mean. I'm personally interested in generally applicable rules, so the fact that it could have worked in this case doesn't help me decide

by blub 7y ago

I think I see what you mean. I'm personally interested in generally applicable rules, so the fact that it could have worked in this case doesn't help me decide what to so when designing another system. In the general case, the Ada behavior of detecting and signaling numeric errors seems to be the only robust choice to me. The alternative could only be acceptable if it's proven that the algorithms fail safe in the presence of tainted data.

kragen 7y ago
I'm interested in generally applicable rules, too; in this case, the generally applicable rule is that there is no generally applicable default robust behavior—the behavior of detecting numeric errors and crashing the system was the worst possible behavior in this context, but in other contexts it would be the best possible behavior. As I explained in https://news.ycombinator.com/item?id=20935662 https://news.ycombinator.com/item?id=20935662, there was no acceptable way to handle such an error if it was detected at runtime. The only acceptable solution is to rigorously verify before runtime that the error is not going to happen, as Rust does with, for example, multithreaded data races. (Well, you can't rule out hardware malfunction, but if you can make it sufficiently improbable you can use the strategy that they did in fact use.) However, default-ignore vs. default-explode isn't a difference between handling errors with IEEE-754-qNaN-style dataflow-tainting and CLU-style nonlocal control flow transfers, as you seem to think. In either case, algorithms that need data that could not be computed successfully will fail to get that data; in either case, you can write the code to either propagate that failure or to attempt to recover from it somehow. In either case, you will not be able to control the spacecraft if you needed that data to control it. The difference I was mentioning, which is not generally applicable, is that in unusual cases like the Ariane 5 disaster, where the erroneous data shouldn't have been computed in the first place because it wasn't needed, qNaN-style error handling clearly results in safer behavior. But, in general, real-time control is not a place where there exists a "robust choice" when your software is broken. If your jet engine control software or your antilock braking software is broken, you're likely to die, and no error handling strategy can prevent that, at least not at the level we're talking about. Biological systems manage this situation by having a melange of somewhat-independent control loops using different algorithms that are constantly fighting for control of the organism; while this does seem to produce systems that are more resilient than engineered systems, it's less efficient and enormously more difficult to develop or understand, and it's the opposite direction from the rigorous certainty you and I favor.

by blub 7y ago

kragen 7y ago

I'm interested in generally applicable rules, too; in this case, the generally applicable rule is that there is no generally applicable default robust behavior—the behavior of detecting numeric errors and crashing the system was the worst possible behavior in this context, but in other contexts it would be the best possible behavior. As I explained in https://news.ycombinator.com/item?id=20935662 https://news.ycombinator.com/item?id=20935662, there was no acceptable way to handle such an error if it was detected at runtime. The only acceptable solution is to rigorously verify before runtime that the error is not going to happen, as Rust does with, for example, multithreaded data races. (Well, you can't rule out hardware malfunction, but if you can make it sufficiently improbable you can use the strategy that they did in fact use.) However, default-ignore vs. default-explode isn't a difference between handling errors with IEEE-754-qNaN-style dataflow-tainting and CLU-style nonlocal control flow transfers, as you seem to think. In either case, algorithms that need data that could not be computed successfully will fail to get that data; in either case, you can write the code to either propagate that failure or to attempt to recover from it somehow. In either case, you will not be able to control the spacecraft if you needed that data to control it. The difference I was mentioning, which is not generally applicable, is that in unusual cases like the Ariane 5 disaster, where the erroneous data shouldn't have been computed in the first place because it wasn't needed, qNaN-style error handling clearly results in safer behavior. But, in general, real-time control is not a place where there exists a "robust choice" when your software is broken. If your jet engine control software or your antilock braking software is broken, you're likely to die, and no error handling strategy can prevent that, at least not at the level we're talking about. Biological systems manage this situation by having a melange of somewhat-independent control loops using different algorithms that are constantly fighting for control of the organism; while this does seem to produce systems that are more resilient than engineered systems, it's less efficient and enormously more difficult to develop or understand, and it's the opposite direction from the rigorous certainty you and I favor.