> In this sense, the sins of the scientific programmer might actually be important
No, you’re turning facts on their head: “fail early”, as described by you, is not a “sin”, on the contrary, it’s a hallmark of good software engineering [1]. Of course a nicely formatted error message is preferred to a coredump, but the end result is similar. But that’s not what (a lot of) bad scientific code does. Instead, it veers into the realm of undefined or unpredictable behaviour by failing to recognise the existence of a problem.
With a core dump one can explore the execution environment at the time of the crash.
A nice compromise is a macro that prints an error message and then calls __builtin_trap().
I don't know how I would read a core dump of an error in a python program to understand where an exception came from.
For example, if you do a simulation with random numbers, and you're doing random()%NUM where NUM > MAX_RAND
The trick is that MAX_RAND varies between platforms cough Visual C cough
random() is often a linear congruential generator (LCG: http://en.wikipedia.org/wiki/Linear_congruential_generator) for speed and simplicity purposes. LCGs are a multiply, an add and a modulus (the modulus is usually implicit from the machine word size). That means their low bits are highly predictable and not random at all.
X(n+1) = (a * X(n) + c) mod m
Assume m is a power of 2 since it's usually implemented via machine word wrap-around. If c is relatively prime with m (in order to fill the whole range of m), then it will be odd. a-1 is normally a multiple of 4 since m is a power of two, so a is odd too.So if X(n) is odd, X(n+1) will be even (o * o+o => o+o => e), and X(n+2) will be odd (o * e+o => e+o => o), and so on, with zero randomness.
So if you're trying to simulate coin flips and use %2, you will get a 1,0,1,0... sequence.
For example, if RAND_MAX is 255, random()%10 equals 0, 1, 2, 3, 4, or 5 26/256 of the time, but 6, 7, 8, or 9 only 25/256 of the time.
That's why good libraries have a 'next(maxValue)' function that is more complex than random()%NUM. For example, see lines 251-268 of http://developer.classpath.org/doc/java/util/Random-source.h....
On Linux, same thing. It even gives the same sequence as the Mac OS version
(it's 100 numbers, "1,0,1,1...0,1," no \n at the end)
./rt | md5sum 7a5a5a0758ca83c95b21906be6052666
rand() is the earliest C random number generator. Its low-order bits (back in the day) went through a predictable sequence, so rand() & 0x1 was a bad source of random bits.
I don't think that rand() was specified so fully as to make this behavior required, but typical implementations exhibited it, so you could not use rand() for any serious work.
random() came after, does not use a LCG, and thus fixed this problem, so you would not see it if your code calls random(), whose man page says:
The difference [between random() and rand()] is that rand() produces a much less
random sequence -- in fact, the low dozen bits generated by rand go through a
cyclic pattern. All of the bits generated by random() are usable. For
example, `random()&01' will produce a random binary value.
Typically, because of this screwup, people use a third-party generator, like the "Mersenne twister".There are other mitigations. For example, Java's RNG uses an LCG, but returns the high 32 bits and uses a 48-bit modulus to counter this weakness.
The built-in random() is not a good source of random numbers for scientific purposes. I've heard too many stories about how people don't think about the random number generator, only to have it bite them in the ass.
Unless you are going to be more specific about when it's not good, this is not generally true.
Scientific purposes often includes Monte Carlo methods, and random() in Python uses a Mersenne Twister - a perfectly good match.
Be careful about calling people idiots.
2 - You're assuming people don't test their generators and just accept blindly whatever result it gives and/or don't compare results of known cases
3 - You say "If you do X you're an idiot" and provide exactly ZERO alternatives to it. Doesn't look very credible
(By "bad" code, I understand code which doesn't meet best practices)
It's very easy to accidentally produce wrong scientific code, partly since scientists are doing research. They use novel mathematical algorithms to solve hard problems, and it's not typically obvious what output is expected. It's not CRUD.
In this sense, the sins of the scientific programmer might actually be important - fragile code which crashes when something is wrong could be considered good - this may help to avoid publishing wrong code.