Slashdot Mirror
'The Code Has Already Been Written'
theodp writes "John D. Cook points out there's a major divide between the way scientists and programmers view the software they write. Scientists see their software as a kind of exoskeleton, an extension of themselves. Programmers, on the other hand, see their software as something they will hand over to someone else, more like building a robot. To a scientist, the software soup's done when they get what they want out of it, while professional programmers give more thought to reproducibility, maintainability, and correctness. So what happens when the twain meet? 'The real tension,' says Cook, 'comes when a piece of research software is suddenly expected to be ready for production. The scientist will say 'the code has already been written' and can't imagine it would take much work, if any, to prepare the software for its new responsibilities. They don't understand how hard it is for an engineer to turn an exoskeleton into a self-sufficient robot.'"
The whole premise is stupid anyway. I've worked with plenty of scientists in national labs that turn out production grade, maintainable code; and programmers who didn't. The core issue is getting people who write code for reuse by others to follow guidelines, regardless of title or profession.
To a scientist, their software is simply a tool, a means to an end. Their results and discoveries are what they really care about. When it comes to reproducing scientific results for verification, it is actually advantageous that another group not use existing software. Another research group using the same faulty software, with the same hidden bugs, will likely come to the same incorrect result.
Productization of software is a completely different exercise. You have to make your software work for a larger crowd on a plethora of devices. You actually have to consider how your software fits into the larger product lifecycle. The key difference here is that you have customers that you need to keep happy.
Kan jeg få en pils, vær så snill?
I work with Monte Carlo code and statistical analysis software. I use CERN's ROOT package for the stats analysis, CERN's GEANT4 for the MC code, and *nix scripting when I need to handle multiple files. Every single piece of code I write is written for a purpose. That purpose is generally to generate data and then analyze it. The only other people who are going to see it? Maybe my supervisor, and, if I'm just in on a contract, maybe the guy who has to work on my code later. But to be blunt, that doesn't matter. All that matters is that I know what's going on.
That being said, sometimes I write software for my own personal use. There, I tend to write more robust code, trying to follow various programming standards. Because I figure, if I write something for myself that turns out to be fairly useful, someone might want to use it, or adapt it. But professionally, all my code needs to do is get out that table or prepare that figure. Is it sloppy? Yes. Does it get the job done? Also yes. Fortunately, not only is my field esoteric, it's also government work, so it's practically a guarantee that my code will never have commercial release.
Cynical Idealist
You can often tell whether someone is "programming as a means to an end (of your own)" versus "programming to build a tool for someone else". For instance, I have experience in the financial industry. Quite a lot of traders see coding as a means to implement their cool new model. Looking at their code, you can often tell. It's as if everything was built to just exactly fulfil the requirement, with no thought to the fact that those requirements might change. But of course, they do change. So you get hacks and workarounds, and cut'n'paste cargo cult code. Kinda like what those Orks in Warhammer 40K might make. And of course the problem with spaghetti code is that if you write it, nobody can ever help you solve problems/improve it. It's the coding equivalent of painting yourself into a corner. There's loads of smart traders out there with an excel spreadsheet that actually is an extension of their personalities (In fact it's their Magnum Opus. Everywhere they go, they try to take this quirky little file with them). Every little hack is something only they can explain (comments, yeah right. Do your body parts have explanatory comments?) and only they can fix if wrong.
On the other hand, you sometimes hire a guy who is a programmer, but knows nothing about the domain. Very good with OO models and that kind, but you have to teach them everything about finance. What's a settlement date, what kinds of options exist, etc. You get what you ask for, because they know how to turn problems into object models, but you have to ask VERY carefully. And teach. Unfortunately, not everyone has time for that, and so you end up with something that still doesn't quite do what it's supposed to.
So you often end up gettings guys who understand the problems, but can't program, programming. And guys who can program, writing the wrong program.
It isn't stupid at all. Lots and lots of scientific software is written by grad students worried about the results, and don't care about the quality of the code itself. Their idea of what is "good code" has no relation to what a programmer who's worked in a production environment would call "good code". And invariably they decide to include libraries from other grad students at other institutions of equal or lesser value. And don't even get me started about documentation...
The whole premise is stupid anyway. I've worked with plenty of scientists in national labs that turn out production grade, maintainable code; and programmers who didn't. The core issue is getting people who write code for reuse by others to follow guidelines, regardless of title or profession.
Because you can point to a few (very few) exceptions does not make the story untrue in the vast majority of cases.
Scientist code is usually a giant JUST-SO story, sufficient to derive the results they need for the task at hand.
They either don't have, or avoid putting in data that will crash the program so limit checking is not necessary.
Crashes are fine if they do nothing more than leave a trail of breadcrumbs sufficient to find the offending line of code.
Output need not be in final form, and any number of repetitive hand manipulations of either the input or the output are fine as long as the researcher does not need to spend more time writing any more elaborate code.
This is perfectly fine. The cabinet maker makes jigs. They are designed for their own shop and no one else has exactly the same saw and exactly the same gluing clamps. When the cabinet maker sells his shop, these jigs become useless. Nobody else knows how to use them.
The scientist who takes the time to do a full fledged, fully documented, maintainable, fail-soft package for analysis of data that is unique to their project and their apparatus is probably not doing very much science, and probably not doing their intended job. That budgets force them into this situation is not unusual.
It happens every day in industry, academics, and research. To hand waive it away by saying you know someone who delivers the full package merely calls into question your own understanding of the meaning of a complete, fully documented, maintainable, transferable, and robust software package.
Sig Battery depleted. Reverting to safe mode.
The problem with that is recognizing what code is going to be reused by others and what isn't.
One way to tell is to ask if this a temporary quick fix for something. If they say yes, assume it will be in production forever.
'The tyrant will always find pretext for his tyranny.' - Aesop's Fables
The issues surrounding transitioning research S/W written by scientists into honest-to-goodness production systems are ones I'm very familiar with.
At my company, a lot of energy has been put into bridging the gap over the years with varying results. I believe that the root cause of the problem is that research S/W is not an end-product; typically for scientists the end-product is a research paper, white paper, proposal paper, etc., for which the S/W is only a tool for getting to the end-product. As soon as the experimental (or proof-of-concept) S/W returns the desired results, the software is considered "done".
In contrast, production S/W is often THE end-product for developers, so a lot more attention is given to robustness, re-usability, etc. All the standard thinking that you want to go into your production S/W.
One big issue for us is that the research S/W is almost always written in Matlab, while the production code is written in C++ and Java. The single largest source of bugs in our systems is porting S/W from Matlab to C++ or Java. (As an aside, please let's not talk about the Matlab 'compiler', nor Octave. -- we've already tried them both, and they're both performance hogs and also create SCM and CM nightmares).
We experimented with requiring that the research S/W be written in C++, but it was a disaster. The scientists couldn't get anything done, and the code was just awful. So, back to Matlab it was.
And, my experience is that people who I have a great deal of respect for, who I consider brilliant in their fields, holding PhD's, etc., have produced the crappiest Matlab code I've ever had the sorrow to read. My favorite instance was the use of these local variable names within a single function of research S/W that was considered "done" (true story):
i
ii
iii
iiii
iiiii
iiiiii
And, of course, little documentation as to the mechanics of the code. And believe me, it gets worse from there. Bear in mind that the code does indeed work for its particular purpose, and may well be ground-breaking in that particular research domain. But "done"? Ready for production? Not without a major porting effort (which is really a re-writing effort). The most mysterious thing to me, though, is that the scientists, for all their intellectual firepower, don't understand that it's a problem.
The solution we've converged on is to require our bizdev to be responsible for funding efforts to rewrite the research code and get it integrated into the product baseline. And, the bizdev types can't proclaim a particular capability "done" (eg., sell it to customers) until they've funded and executed those efforts. It took years of education to get to this point, but things are moving along much better then before.
In the course of every project, it will become necessary to shoot the scientists and begin production.
Exactly this! It is not about the education of the people writing the code. It's about the purpose for which it is written. I've done it all.
As a scientist most software that I write is geared to solving the problem at hand, nothing more. Sometimes this can be 10.000 lines of C++ code, at other times a short python script or 10. Each time, the code serves as a sort of automation for something that needs to be done anyway (I could attempt to compute the simulation result myself, by hand on pen and paper, you know... if I don't die of old age first ;) ). Often, not a single thought goes into how to make this stuff reusable, robust or more generic. It works on the one machine it is ever going to run on and very likely nowhere else, because it does not matter. What matters is the program output, not the program itself.
As a software developer I have to think differently. Software gets compiled, packaged and deployed elsewhere. It must run out of the box, never crash, give useful error message and recover cleanly if something bad happened. And amidst all this effort, there's a tiny bit of code hidden somewhere doing the actual work. All that matters is that the program behaves correctly, no matter what the concrete output is. I might not even be expected to understand what the output actually means - it's not my primary concern.
See the difference?
http://www.moonlight3d.eu/
The bulk of scientific research is done by grad students (or others like them with various kinds of scholarships). The professors whose name is at the end of paper's author list guide and oversee what is done, but don't have time for the daily grind of research. Their main job is to teach and get funding.
I always like the Numerical recipes quote: Scientists solve next years problem on last years computer. Computer programmers solve last years problem on next years computer.
I've lived on both sides of this divide but mainly on the scientific side. I become apoplectic with software engineers who just don't vest themselves in the science. The perpetually want a set of requirements. And they get upset if a new requirement is added later. I see software as a way to explore a space. Model it. Determine what more modeling is needed. You are constantly trying to do something that usually is beyond what is computationally possible so you have to figure out what approximation is going to work. What has to be done at full scale and what can be done at lower resolution. Mock up stuff.
The engineers who don't see it as a process just are impediments. Scientists want lots of simple things fast then see what is working and add new simple extensions. They don't want to wait 4 months for some delivered code based on specs it took 2 months to write.
Hence scientist tend to write their own code.
Some drink at the fountain of knowledge. Others just gargle.
That's a great post, of the kind that saves me a lot of typing. You covered the first-order considerations brilliantly.
What you missed was technical debt blindness, which has been around since forever. Books I read around the time of the Mythical Man Month talked a lot about maintenance syndrome: that the original development team would be regarded as brilliant for producing working functionality at tremendous speed (undocumented, with no error handling for edge cases), then the first maintenance team would all be fired as underachievers for adding hardly any new functionality in the first year or two.
Turns out it's hard to erect a machine shop over top of adobe mud brick construction without adding some reinforcement to the structure, which usually takes a lot longer than the entire original edifice.
You can instead take a wrecking ball to the first iteration, but this rarely works out as well as hoped. You end up with far more ambitious adobe mud construction built with a whole new generation of unproven tools. At some point you have to bite the bullet and ferment what you began with.
People hide debt blindness behind widely divergent construals of simplicity, where "simple" usually turns out to be a euphemism for any decision that sidesteps paying down debt in the short term.
For professional software engineers, there is one true simplicity to rule them all: generativity and compositionality. Can you build the next layer on top with any hope of having it work and able to support an ongoing stack? For us, it's a long term game of pass the baton. For everyone else (management, scientists) the endgame is to cash out, and take credit elsewhere (e.g. publication biography).
Unfortunately, a citation is not a formal linkage that the compiler either accepts or rejects. By the standards of compositionality, citation is payment in dubious coin. Citation is not falsifiable. Scientists still count their citations even when they come from papers that are full of crap, peer review notwithstanding. For a professional software engineer, when you start instantiating objects from one library inside an abstract expression template library, you come face to face with compositionality in a way that few scientists can even imagine, having weened at the outrage of being improperly cited.
Technical debt blindness on the part of management quickly turns a software engineering shop into a highly non-linear fiasco. We've all seen this.
Somehow this game works out better (for the participants) when played by bankers with leverage debt. But now it's my turn to pass the baton, since that deserves a whole lot more typing and I've done my bit.
"I'm working on commercializing NASA software and this couldn't be more true."
I don't think is just NASA. Wasn't an IBM engineer the one that said: X time for a valid prototype; 10*X for an in-house product; 100*X for a commercial packaged application?
Given that what a scientist does for her calculations is in the "valid prototype" stage (why the hell would she be interested in anything else?) you can do the math about how much it will cost to put it on a shelf, so to say.