Slashdot Mirror
The Law of Leaky Abstractions
Joel Spolsky has a nice essay on the leaky abstractions which underlie all high-level programming. Good reading, even for non-programmers.
← Back to Stories (view on slashdot.org)
← Back to Stories (view on slashdot.org)
Although I used to program as a hobby, my eyes bugged out when I saw this article. It's actually quite interesting; I finally realize why the hell people program in lower level languages.
One point that I think could be addressed is backward compatibilty. I really know nothing about this, but don't the versions of the abstractions have to be fairly compatible with each other, especially on a large, distributed system? This extra abstraction of an abstraction has to be orders of magnitude more leaky. The best example I can think of is Windows.
I'm of the idea that the whole premise that high-level tools and high level abstraction coupled with encasulation are the biggest bane of the software industry. We have these high level tools which most programmers really don't understand and are taught that they don't need to understand in order to build these sophisticated products.
Yet, when something goes wrong with the underlying technology they are unable to properly fix their product because all they know is some basic java or VB and they don't understand anything about sockets or big-endian/little endian byte alignment issues. It's no wonder todays software is huge and slow and doesn't work as advertised.
The one shining example of this is FreeBSD, which is based totally on low level C programs and they stress using legacy program methodologies in place of the fancy schmancy new ones which are faulty. The proof is in the pudding, as they say, when you look at the speed and quality if FreeBSD, as opposed to some of the slow ponderous OS's like Windows XP or Mac OSX.
Warmest regards,
--Jack
Wagner LLC Consulting Co. - Getting it right the first time
Well I wouldn't say that it's reliable "because there are timeouts". AAMOF, timeouts just compicate things. So you timeout waiting for packet N, you request a resend of it, and in the interim, guess what, packet N shows up, now you have two N's. Your code is now more complex in having to deal with this situation. Timeouts are just another parameter used adjust the behaviour of the algorithms that control the protocol. Getting deterministic results from an undeterministic foundation involves making observations, accepting some compromises, making some simplifying assumptions, and then writing code that takes all those things into account to come up with something that usually works.
Is our own bodies.
I'm studying to be a bioinformatics guy with the university of melbourne and have just had the misfortune of looking into the enzymatic reactions that control oxygen based metabolism in the human body.
I tried to do a worst case complexity analysis and gave up about half way through the krebs cycle.
When you think about it, most of basic science, some religeon and all of medicine has been about removing layers of abstraction to try and fix things when they go wrong.
...to start with, or at least be competent with, the basics.
Any good programmer I've ever known started with the lower level stuff and was successful for this reason. Or at least plowed hard into the lower level stuff and learned it well when the time came, but the first scenario is preferable.
Throwing dreamweaver in some HTML kiddie's lap, as much as I love dreamweaver, is not going to get you a reliable Internet DB app.
vk.
Maybe someone out there prefers to program without any abstraction layers at all, but they inherit so much complexity that it will be impossible for them to deliver a meaningful product in a reasonable time.
FIne it's relaibale becasue of acks, timeouts, adaptive re-transmit timeouts that take statistical averages of RTT times, exponential back-off and slow start, window acks which keep track of what bytes are received, etc.
So in your case of timing out N, re-tx'ing N, and then getting the repsonse to the first N back after sending the second N, you do two things:
1) Good! You got yr packet!
2) keep track of how many bytes you have received thsu far (TCP is not sending messages, it is sending a stream)
3) when you get the response from your second request, discard it, becuase you already received those bytes from the stream.
4) since you timed out, DON'T use the Round TRip Time for that reponse: slow down your expected RTT time, and THEN start measuring.
And guess what? If I unplug the NIC of the other machine, there is no reliable way of transmitting that data (assuming your destination machine isn't dual homed)- so I keep streaming bytes to a TCP socket and I don't find out my peer is gone for approx. 2 minutes.
WOW. There's nothing reliable about that boundary condition!
my point is TCP is reliable ENOUGH. But I wouldn't equate it with a Maytag warranty. It is not a panacea. Infact, for a closed homogenous network I wouldn't even consider it the best option. But if the boundary conditions fall within the acceptible fudge range (remember Real Time human grade systems are not 100% reliable, only 99.99999% and much of that is achieved through redundancy) your leaks are ok.
In the future, I would want to not be isolated from my friends in the Space Station.
The market rewards abstractions because they help create high level tools that get products on the market faster. Classic case in point is WordPerfect. They couldn't get their early assembler-based product out on a competitive schedule with Word or other C based programs.
Sure, the author points out a few examples of leaky abstractions. But his conclusion seems to be that you always will have to know what is behind the abstraction.
I don't think that's true. It depends on how the abstraction is defined, what it claims to be.
You can use TCP without knowing how the internals work, and assume that all data will be reliably delivered, _unless_ the connection is broken. That is a better abstraction.
And the virtual memory abstraction doesn't say that all memory accesses is guaranteed to take the same amount of time, so I don't consider it to be leaky.
So I don't entirely agree with the author's conclusions.
Proper abstractions avoid unintended side-effects by presenting a clean view of the intent and function of a given interface, and not just a collection of methods or structures.
When I read what Joel wrote about "leaky abstractions" i saw a peice complaining about "unintended side-effects". I don't think the problem is with abstractions themselves, but rather the implementation.
He lists some examples:
1. TCP - This is a common one. Not only does TCP itself have peculiar behavior in less than ideal conditions, but it is also interfaced with via sockets, which compound the problem with an overly complex API.
If you were to improve on this and present a clean reliable stream transport abstraction is would likely have a simple connection establishment interface and some simple read/write functionality. Errors would be propagated up to a user via exceptions or event handlers. But the point I want to make is that This problem can be solved with a cleaner abstraction.
2. SQL - This example is a straw man. The problem with SQL is not the abstraction it provides, but the complexity of dealing with unknown table sizes when you are trying to write fast generic queries. There is no way to ensure that a query runs fastest on all systems. Every system and environment is going to have different amounts and types of data. The amount of data in a table, the way it is indexed, and the relationship between records is what determines a queries speed. There will always be manual performance tweaking of truly complex SQL simply because every scenario is different and the best solution will vary.
3. C++ string classes. I think this is another straw man. Templates and pointers in C++ are hard. That is all there is too it. Most Visual Basic only coders will not be able to wrap their minds around the logic that is required to write complex c++ template code. No matter how good the abstractions get in C++, you will always have pointers, templates, and complexity. Sorry Joel, your VB coders are going to have to avoid c++ forever. There is simply no way around it. This abstraction was never meant to make things simple enough for Joe Programmer, but rather to provide an extensible, flexible tool for the programmer to use when dealing with string data. Most of the time this is simpler, sometimes it is more complex (try writing your own derived string class - there are a number of required constructors you must implement which are far from obvious) but the end result is that you have a flexible tool, not a leaky abstraction.
There are some other examples, but you see the point. I think Joel has a good idea brewing regarding abstractions, complexity, and managing dependencies and unintended side-effects, but I do not think the problem is anywhere near as clear cut as he presents. As a discipline software engineering has a horrible track record of implementing arcane and overly complex abstractions for network programming (sockets and XTI) generic programming (templates, ref counting, custom allocators) and even operating systems API's (POSIX).
Until we can leave behind all of the cruft and failed experiments of the past, start new with complete and simple abstractions that do not mask behavior, but rather recognize it and provide a mechansim to handle it gracefully, we will run into these problems.
Luckily, such problems are fixable - just write the code. If joel were right and complex abstractions were fundamentally flawed, that would be a dark picture indeed for the future of software engineering (it is only going to grow ever more complex from here kids - make no mistake about it).
We have examples of massively complex systems that work very reliably day-in and day-out. Jet airplanes, for one; the national communications infrastructure, for another. Airplanes are, on the whole, amazingly reliable. The communications infrastructure, on the other hand, suffers numerous small faults, but they're quickly corrected and we go on. Both have some obvious leaky abstractions.
The argument works out to be pessimism, pure and simple -- and unwarrented pessimism to boot. If it were true that things were all that bad, programmers would all _need_ to understand, in gruesome detail, the microarchitectures they're coding to, how instructions are executed, the full intricacies of the compiler, etc. All of these are leaky abstractions from time to time. They'd also need to understand every line of libc, the entire design of X11 top to bottom, and how their disk device driver works. For almost everyone, this simply isn't true. How many web designers, or even communications applications writers, know -- to the specification level -- how TCP/IP works? How many non-commo programmers?
The point is that sometimes you need to know a _little bit_ about the place where the abstraction can leak. You don't need to know the lower layer exhaustively. A truly competant C programmer may need to know a bit about the architecture of their platform (or not -- it's better to write portable code) but they surely do not need to be a competant assembly programmer. A competant web designer may need to know something about HTML, but not the full intricacies of it. And so forth.
Yes, the abstractions leak. Sometimes you get around this by having one person who knows the lower layer inside and out. Sometimes you delve down into the abstraction yourself. And sometimes, you say that, if the form fails because it needs JavaScript and the user turned off JavaScript, it's the user's fault and mandate JavaScript be turned on -- in fact, a _good_ high-level tool would generate defensive code to put a message on the user's screen telling them that, in the absence of JavaScript, things will fail (i.e. the tool itself can save the programmer from the leaky abstraction).
What Ian Malcolm says, when you boil it all down, is that complex systems simply can't work in a sustained fashion. We have numerous examples which disprove the theory. That doesn't mean that we don't need to worry about failure cases, it means we overengineer and build in failsafes and error-correcting logic and so forth. What Joel Spolsky says is that you can't abstract away complexity because the abstractions leak. Again, there are numerous examples where we've done exactly that, and the abstraction has performed perfectly adequately for the vast majority of users. Someone needs to understand the complex part and maintain the abstraction -- the rest of us can get on with what we're doing, which may be just as complex, one layer up. We can, and do, stand on the shoulders of giants all the time -- we don't need to fully understand the giants to make use of their work.
I've heard a lot of people say that they can't believe how many homes, schools, and other buildings were destroyed by the huge thunderstorms that hit the states this past weekend, or that many people died. Hello, we haven't yet figured out how to control everything! American (middle to upper-class) life is a leaky abstaction. We find this out when we have a hard time coping with natural things that shake up our perfect (abstacted) world. That is what we all need to understand.
Is everyone as scientifically and technically literate and educated as they should be? Of course not. Are those involved in liberal arts education the worst culprits? Of course not. A liberal arts education most certainly includes a healthy dose of mathematics and science, particularly of the kind you highlight, namely higher-level abstractions.
What is terrifying about your post is the final paragraph where you imply that an engineering background is better for those who want to practice ``politics, humanities, governance, management''. Scientists and engineers are like any other group. They can have narrowly-focused educations, minds, and perspectives. The fields that you mention are complex, and require study, background, consideration, and debate of their own.
While you are slamming ``liberal arts'' -- a term you seem not to understand -- you highlight the need for it. Liberal arts does not imply a non-scientific, non-technological education. It implies a broad education, including science, mathematics, and engineering along with the ``traditional'' topics of history, literature, languages, politics, economics, and arts. For politics, governance, and management, I want people who are conversant in all of those topics.
Don't fix what ain't broke. You would have saved even more time (And bugs!) if you just let the Fortran be. Seriously.
While I usually like Joel's work, I'm pissed about the random jab at C++. For those he didn't read the article, he says something along the lines of
"A lot of the stuff the C++ committe added to the language was to support a string class. Why didn't they just add a built-in string type?"
It's good that a string class wasn't added, because that lead to templates being added! And templates are the greatest thing, ever!
The comment shows a total lack of understanding of post-template, modern C++. People are free not to like C++ (or aspects of it) and to disagree with me about templates, of course, and in that case I'm fine with them taking stabs at it. But I get peeved when people who have just given the language a cursory glance try to fault it. If you haven't used stuff like Loki or Boost, or taken a look at some of the fascinating new design techniques that C++ has enabled, then you're in no place to comment about the language. At least read something like the newer editions of D&E or "The C++ Programming Language" then read "Modern C++" before spouting off.
PS> Of course, I'm not accusing the author of being unknowledgable about C++ or anything of the sort. I'm just saying that this particular comment sounded rather n00b'ish, so to speak.
A deep unwavering belief is a sure sign you're missing something...
The problem that Joel talks about is not really a problem with abstraction, it is a problem with teamwork. When I program I simply can not do all of it myself for all but trivial projects. By everything I mean write the compilers, write the OS etc. Instead I must rely on other programmers to write large portions of the code that I run. Whether it is the guy across the hall who wrote the search contact Stored Procedure in SQL or a programmer at microsoft writing a windows Disk IO function, I am relient on their code working as I think it should. This is the problem with teamwork but there is no other solution to programming modern applications. Martin
Maybe I'm an old fashioned has-been but people doign software development should understand the fundamentals of how computers work. That means that they should understand things like memor management, they should understand what a pointer is, they should undertsand about how tight loops versus unrolled loops might affect the performance of the caches on their system. I meet so many "programmers" that have no understanding that there are architectural constraints on what they can and can't do. Software runs on hardware. If you're going to write software and treat the hardware as a black box, you're not going to write it as well, or as efficiently as you could be doing it.
Now I always consider performance when designing/writing code, but programmers are WAY more expensive than hardware, so eeking out performance can often be a wasted effort. Everyone knows that C will smoke Java in most operations, but having its so hard to manage at the enterprise level that you are much better taking the 50%+ performance hit and writing in a "leaky" language.
This is not the greatest sig in the world, this is just a tribute.
Looks like he just discovered and renamed the basic idea that "all models are incomplete". Any scientist could tell you that one! I remember a quote that goes something like this: The greatest scientific accomplishment of the 19th century was the discovery that everything could be described by equations. The greatest scientific accomplishment of the 20th century is that nothing can be described by equations.
That's all an abstraction is: a model. Just like Newtonian physics, supply and demand under perfect competition, and every other hard or soft scientific model. Supply and demand breaks down at the low end (you can't be a market participant if you haven't eaten in a month) and the high end (if you are very wealthy, you can change the very rules of the game). Actually, supply and demand breaks down in many ways, all the time. Physics breaks down at the very large or very small scales. Planetary orbits have wobbles that can only be explained by more complex theories. Etc.
No one should pretend that the models are complete. Or even pretend that complete models are possible. However, the models help you understand. They help you find better solutions (patterns) to problems. They help you discuss and comprehend and write about a problem. They allow you to focus on invariants (and even invariants break down).
All models are imperfect. It's good that computer science folks can understand this, however, I don't think Joel should use a term like "leaky abstraction". Calling it that implies the existence of "unleaky abstraction", which is impossible. These are all just "abstractions" and the leaks are unavoidable.
Example: if I unplug the computer and drop it out of a window, the software will fail. That's a leak, isn't it? Think of how you would address that in your model: maybe another computer watches this one so it can take over if it dies..etc..more complexity, more abstractions, more leaks....
He also points out that, basically, computer science isn't exempt from the complexity, specialization, and growing body of understanding that accompanies every scientific field. Yeah, these days you have to know quite a bit of stuff about every part of a computer system in order to write truly reliable programs and understand what they are doing. And it will only get more complex as time goes on.
But what else can we do, go back to the Apple II? (actually that's not a bad idea. That was the most reliable machine I've ever owned!)
You don't, and in fact can't, deal with page faults in your Java program. Nonetheless, your java program will suffer a performance hit when it page faults. Thats a leaky abstraction.
While you are slamming ``liberal arts'' -- a term you seem not to understand -- you highlight the need for it. Liberal arts does not imply a non-scientific, non-technological education. It implies a broad education, including science, mathematics, and engineering along with the ``traditional'' topics of history, literature, languages, politics, economics, and arts. For politics, governance, and management, I want people who are conversant in all of those topics.
Unfortunately, the subjects you list have all grown to the point that no human can obtain even a BASIC understanding of all of them before he's too old to have a useful carreer left.
It was once possible to be a "Rennisance Man" - a master of ALL the sciences and arts reduced to teachability. No more. It's just too bloody large. (I say this as someone who attended a univerdity that claims to try to produce such people - centuries after the last of them is dead. B-) )
Unfortunately, "Liberal Arts" schools have, over much of the last century, been filled with the mathematically and technically illiterate - both because the students without the necessary skills gravitated there, and because the faculties themselves were so disabled, and in turn disparaged the skills they were incompetent to teach.
The engineering/scientific/biologic/technical cirriculum had constant feedback from the real world about what was true and what was false. But the "Arts Schools" taught classes where what was "right" was ONLY a matter of opinion - and grades solely a measure of how well you could regurgitate your Prof's pet bonnet-bees. (This DESPITE the fact that SOME of these theories could be TESTED - if only the academics understood, and/or believed in, things like the scientific method, statistics, and sampling methods.)
Yes the "Social 'Sciences'" are hard. But the bulk of their credentialed practitioners used this as an excuse to drop "science" from their methodologies. (This despite that fact that mathematics departments were generally part of the art, rather than the engineering, side of the school organization.)
I've been out of academia for a while now. I can hope that things have improved, as you seem to claim. But I have not personally seen any sign of such from the outside (other than your claim).
In my school days, too, many students on the Arts side of the wall knew tech, math, and the like. (Students are generally young, and still hunting for their muse.) But they would generally transfer out to some field more conducive to clear thought, drop out to use it in the real world, or (if they stayed in LS&A) suppress it or flunk out.
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way
I don't see how you write a can get a String class function to be called using only char*'s as arguments. Thus the abstraction fails.
Hello Cruel World
The subject of leaky abstractions applies to novice users as well.
I've felt for a long time that people are taught about computers the wrong way, and this article clarifies why this is true.
People are taught less and less about what the computer actually does, and instead focus on things like the desktop analogy, and task oriented training. The user must then remember all these seemingly strange things computers do that don't follow the abstraction they were taught. This makes them seem difficult and incomprehensable.
The problems created by abstractions intended for users can simply be solved with more complicated software that better models the analogy that the users are taught. Unfortunately, the opposite is probably true for programmers.
-Zaphod
Please tell me what you think of this - I would honestly like to know.
I've worked in a way similar to you, and I might still if it were as mindlessly simple to write assembly language programs under Windows as it was back in the day of smaller machines (i.e. no linker, no ugly DLL calling conventions, smaller instruction set, etc.). In addition to being fun, I agree in that assembly language is very useful when you need to develop your own abstractions that are very different from other languages, but it's a fine line. First, you have to really gain something substantial, not just a few microseconds of execution time and an executable that's ten kilobytes smaller. And second, sometimes you *think* you're developing a simpler abstraction, but by the time you're done you really haven't gained anything. It's like the classic newbie mistake of thinking that it's trivial to write a faster memcpy.
These days, I prefer to work the opposite way in these situations. Rather than writing directly in assembly, I try to come up with a workable abstraction. Then I write a simple interpreter for that abstraction in as high a level language as I can (e.g. Lisp, Prolog). Then I work on ways of mechanically optimizing that symbolic representation, and eventually generate code (whether for a virtual machine or an existing assembly language). This is the best of both worlds: You get your own abstraction, you can work with assembly language, but you can mechanically handle the niggling details. If I come up with an optimization, then I can implement it, re-convert my symbolic code, and there it is. This assumes you're comfortable with the kind of programming promoted in books the _Structure and Interpretation of Computer Programs_ (maybe the best programming book ever written). To some extent, this is what you are doing with your macros, but you're working on a much lower level.
The abstraction is a reliable byte stream, which of course isn't really possible due to phenomena that can only be affected by interfaces beneath TCP. A leak that's documented is still a leak.