Slashdot Mirror
Debugging in Plain English?
sameerdesai writes "CNN is carrying a story about Researchers from Carnegie Melon: Myers and a graduate student, Andrew Ko, have developed a debugging program that lets users ask questions about computer errors in plain English: Why didn't a program behave as expected? I guess with recent exploits and bugs that were found this will soon be a hot research topic or tool in the market." We recently did a story about revolutionary debugging techniques; the researchers' website has some papers and other information.
Untill I can have a full conversation with a computer (a la the Turing effect, not the limited versions that Alice et. al. can accomplish) I'll be happy with source code, thank you very much. It's just another layer blocking me from the code anyway (read In the Beginning... lately?).
All these talks of "revolutionary" debugging techniques bother me a little. There's only one debugging technique, and that's the debugger's skill and experience. Debuggers, traces, logs and other printf()s and LEDs flashing are just tools.
Andrew Ko's invention is just another tool. It won't do the debugging for you. Just like modern cars have diagnostic computers, but somehow it appears you still have to fork off $30/hr for the workmanship to get it fixed...
"A door is what a dog is perpetually on the wrong side of" - Ogden Nash
Article says adding Whyline to java makes it 10x more complex. Seems to me like just another example of Computer Science grad students trying to justify their existence.
How useful will this be when it responds to a simple question with a simple answer?
me: why did the program leak 1GB of memory then segfault?
computer: because you don't know how to program, you idiot!
There seems to be a fundamental logic problem with this kind of debugger. It seems to be more of an 'interpretive debugger' in which it would need to know what you're trying to do in the first place. Thus, it would have to know what the code -should- look like in order to tell you what the problem is.
In essence, for there to be a "English debugger" (one that speaks more english than current debuggers, that is), it would essentially need to know how to program itself on top of being able to follow the flow of code and find where it breaks, so as to be able to tell you precisely what the problem is.
Sounds a bit fictional to me.
~/ssh slashdot.org ssh: connect to host slashdot.org port 22: too many beers
When I think about some of the bugs I have found (and coded), the Whyline approach seems very far-fetched. The degree of self-awareness (introspection) required by something like this makes it seem like the program would be able to avoid the trap in the first place. It would require the "analyst" or "observer" module to observe not only a stack trace and PC trail, but also would require the module to understand what is supposed to occur.
I don't expect this early research tool to catch all of these, but I'd like to hear the researchers' response on how their system might (after years of development) answer questions about some of these bugs:
- Why did the Mars Pathfinder software deadlock (priority inversion)
- Why did the Mars Polar Lander crash (improper state management)
- Why did the Ariane 5 blow up (arithmetic overflow in a register)
- Why did the Patriot missiles miss in the 1991 Gulf War (accumulated time error)
- Why did a radiation therapy machine zap patients with the wrong doses (inconsistent state between GUI display and internal software state)
I'm sure there are some others on comp.risks and elsewhere.
Another point: this approach is still "just" a testing tool. In other words, it can only find errors on paths that have actually been taken in tests, which means the testing program must cover enough cases to generate the runtime errors in the first place. In all of the above cases, it was the testing program that permitted the bugs to be fielded.
In the area of error messages, we need a better balance between "understandable" and "informative". Today, it's usually one extreme or the other. Either error messages are too "friendly" to the point that they're rather meaningless (as in the notorious "Web site not found"), or on the other extreme, they're so "complicated" that the novice user has no idea what happened (as in "an xxx exception occurred at xxx, here is the stack trace, register states, and many many more confusing numbers").
It would be nice if we (as a society that happens to use computers) could adopt a guide that gives the technically-savvy enough info to go on, but still helps the novices understand what went wrong in terms they can understand too. A balanced approach.
Such an error reporting scheme, if successful, could almost certainly NOT be designed by marketing weenies nor by geeks alone, but through mutual collaboration and willingness to compromise on both sides. Hmm, no wonder it's so challenging.
I'm an animal lover -- they're delicious!
Programmer: "Why didn't it work?"
Computer: "How should I know, I just do what I'm told."
"I use a Mac because I'm just better than you are."
Unfortunately it's because COBOL is excrutiatingly bad at expressing the sorts of things programmers need to express. It's not so hot for expressing business information either.
I am reminded of the attempts to cross horses with zebras -- it worked, but the hybrid inherited the worst traits of each animal.
DNA just wants to be free...
This reminds me of back in approx 1985 or so, someone "invented" a human language programming environment called "The Last One" or something like that. This would supposedly make it simple to write programs without having to learn C etc. Some programmers quaked in their boots. However the real issue with programming is learning the contructs, not the language (ie. if you understand what a linked list is, then writing one in C vs Pascal is pretty simple). Anyone that thinks that programming in English is easier is seriously misunderstanding programming. The ultimate test is to look at the languages that have survived: The more "human readable" languages like COBOL have not survived, but the more cryptic ones like C have. The big "killer app" for making programming simple for the non-programmer was the spreadsheet and that's hardly a natural language.
Now debugging is pretty much the same deal. Verbose English debugging interfaces might make it simple to learn to do very basic debugging, but once you get into things a bit deeper (and get more experienced), English becomes a huge liability and you'll be wishing for more concise and expressive languages.
Engineering is the art of compromise.
With these "natural language" languages, I always wonder- is English the only natural language out there? Doesn't look like it from the world around me.
Then why all those "natural language" thins are in English?
Seems to me that is easier to learn a programming language then to learn English for a non-native English speaker: much less ambivalence.
xslf.com
I was going to write up a cardboardprogrammer.com site with a flowchart with 20 questions to ask about a bug to clarify your thoughts. [I suppose both linux and windows can have a REBOOT as the first directive and what was the difference between now and when it was last working as the second, are you using the latest version of the code as the third has anyone else been working on these files the fourth.. well you get the idea, [it is like shooting fish in a barrel, but I have never seen it the full flowchart] that was in the dotcom era, when even a tiny good small idea like this was in someones mind a possible next big thing].
Be Free: Free Software Tuition
In the products I'm involved with I often get stupid reports from the field of the form "framing error causes unit to reset". When I get one of these, first thing I do is get back to the user and figure out exactly what they saw and what heppened withouth them trying to figure out the symptoms. In the "framing error" problem, what was really happening was that a power glitch was being caused when the RS232 cable was attched (because of bad grounding). This caused a reset. However, the user was a "super user" who knew bad things happen to serial data when you plug/unplug cables. One of the buzzwords he knew about was a framing error. So he "half solved" the problem by saying that a framing error caused the problem.
There is a big difference between observing and fixing problems. QA is about observing, not fixing, problems. It is better to provide a good way for users to make accurate error reports (eg. backtrace/log/whatever) than try have them try explain what went wrong.
Engineering is the art of compromise.
To go from 1 to 2 is the hard part. But, you can use tools to instrument the problem and determine where the problem may be.
Fight Spammers!