Cannot Measure Productivity
martinfowler.com
martinfowler.com
Software is a form of literacy - and we measure literacy completely differently. In fact we measure it like we measure science - you are not a scientist unless other scientists agree you are, and you are not a coder unless other coders say you are.
What Fowler wants to measure is not the top echelons of productivity but the lower bounds - presumably to winnow out the unproductive ones.
But that is not how we conduct ourselves in literacy or science. We educate and train people for a very long time, so that the lower bound of productivity is still going to add value to human society - and the upper bounds are limitless.
What Fowler is asking for is a profession.
Science requires one thing: making and testing falsifiable hypotheses. A priest is able to determine whether or not you are doing that. If anything, it's philosophers who decide what science is, e.g. Karl Popper.
So, I too consider that agreement from others isn't a prerequisite for being a scientist. I also agree that "making and testing falsifiable hypotheses" definitely qualifies as doing science.
However, perhaps that's not the only way to do science. In general, there is that whole set of criticisms on the limits of falsifiability (with Kuhn et al). In particular, I'm thinking of cases where arguably the technology isn't sufficiently advanced to perform the measurements necessary to directly test the hypotheses (e.g., how quantum physics progressed). Arguably those doing all the thought experiments, modeling, thinking through consequences of those hypotheses and comparing with what they could measure were doing science -though those weren't falsifiable hypotheses at the time.
So what I'm saying is that your requisite is sufficient but perhaps isn't necessary either.
"Scientist" implies a certain amount of knowledge, training, discipline, etc. I'm not implying that every scientist needs to have undergone academic training - there are other ways - but merely doing a scientific experiment is not enough to call yourself a scientist.
A scientist is one who "does science" with some knowledge, consistency and perseverance.
> In a more restricted sense, a scientist is an individual who uses the scientific method. [...] This article focuses on the more restricted use of the word.
If we're going to throw definitions around, how about dictionary.com: http://dictionary.reference.com/browse/scientist?s=t
> an expert in science, especially one of the physical or natural sciences.
If being a scientist is determined by the consensus of one's peers, it seems like it makes sense to accept an article defining what scientists are that is written as a consensus opinion.
But anyway, if you think it's wrong, why don't you edit it?
The open questions are 1) whether it's possible to do that, and 2) whether it's possible to do that without inadvertently imposing an upper bound. Common wisdom among software developers is that it is not possible and even if it was it would impose a very tight upper bound. But as far as I know, that common wisdom is just guess work. No one has found a good solution yet, but there's also no proof that there is not one. That's why so many people continue to try.
What parts of what he writes do you disagree with?
I suppose he ends with throwing his hands in the air and saying we cannot do it. Well actually we can achieve a measure of literacy, just not perhaps fluency.
We know if a child can read and read well. It's possible to take their written work and assess or mark it. We can (and do in code reviews) do the same for code written by adults.
The problem he is stating is IMO in two parts - measuring basic competence (is a person of net benefit, can they write competent code, are they literate?) which is possible, and how good are they compared to their peers, which is much more subjective and based on taste as much as anything.
So I agree we should not try to measure productivity, I disagree that that's the end of the conversation, mostly because I disagree that productivity is one thing - I see it as more Hygiene/Motivator
Not everybody needs the approval of an external entity. I couldn't care less what other coders think of me. What I ship speaks for itself. What I write speaks for itself. What studies scientists publish speak for themselves. The rest is politics.
Coders must get jobs to continue coding if they are not independently wealthy. Bias abounds in this case, as other coders with influence must vouch for you, and...in this case your projects can influence their opinion of you.
We are social creatures: we don't live as hermits (most of us anyways).
In the days before the Web, I created two good-sized adventure games without ever having interacted with another programmer (save maybe 2 book authors). Nobody vetted me. I just made and released them. The lone wolf programmer has always been a thing.
This is why I use the term software literacy. My son is learning to read. He can write is own name and letters, read some words phonetically. All of those things are necessary but not sufficient. But he is not (yet :-) literate. Will it be at ten words ? A hundred? A thousand? Those are silly arbitrary cut offs.
Anyone of us here can tell the difference between literate and illiterate because we have passed through that gateway.
The same goes for scientist or programmer.
But actually trying to write down the exact definition, the point one passes from being able to write a line of executable code and becomes a real programmer? Becomes software literate? Can't do it. Which is also why you can't measure productivity (plus all the reasons Robert Austin has)
100 years from now, perhaps there will be "Licensed Professional Programmer" certifications. Until then, you're a professional programmer if someone pays you. Even if that someone is yourself.
Really. No reality check? Spinning a wheel doesn't cause anything in particular, unless it meets the road.
Now I am not saying what you are saying is totally false. The truth is more nuanced, or say, more qualified.
The qualification comes from achieving something in the real world. If you can achieve something by talking to someone instead of writing code to work around it, both are equally valid.
I think of it like this. A pure function does not do any real work. The real work comes from the side effect it causes, the global variable it sets, the file it writes to, the program it talks to on the other end.
So yeah. You can cocoon yourself saying lalalalala, but you don't want to be the thread that spawned, did something to its local variables and exited. Whats the point of such a life anyways?
The context of my statement is in regards to self identification vs community identification. I write code, I release apps and I get paid. Why do I need a community to call me a coder before I can be considered one? What other people think of me, what image I put forth, etc is all marketing and politics. The only real evidence of me as the coder is what I ship. Does it scale? Does it have bugs? Does it work? That's the reality check, not whether Steve from SuperFrog Backup Solutions saw my code on github and thinks I write an elegant monad.
If you can achieve something by talking to someone instead of writing code to work around it, both are equally valid.
Nothing I wrote disputes that.
So yeah. You can cocoon yourself saying lalalalala, but you don't want to be the thread that spawned, did something to its local variables and exited. Whats the point of such a life anyways?
Consider this: a man spends a lifetime writing novels. He thinks of himself as a writer, he identifies himself as a writer and he introduces himself as a writer. He never publishes, but he's always producing. One day, his house burns down, killing him and destroying everything he's ever written. Do you think his life was wasted? Do you think he thought his life was wasted?
Still, I can't help cringe at some of the ideas posited -- for one, the idea that more features == a better product.
Or that profit is a measure of engineering productivity. As if it's that uncommon for great products to be badly marketed.
However I do think that (free) market success is a reasonable measure of value / worth - backed up by my favourite quote for which I can find no author:
"Many books are unfairly forgotten, but none are unfairly remembered."
Feynman had some interesting thoughts on minimal computation that sort of paralleled Shannon's information complexity. As you know Shannon was interested in absolute limits to the amount of information in a channel and Feynman was more about the amount of computation per joule of energy. But the essence is the same, programs are a process that use energy to either transform information or to comprehend & act on information so 'efficiency' at one level is the amount of transformation/action you get per kW and "productivity" is the first derivative of figuring out how long it takes to go from need to production.
It has been clear for years that you can produce inefficient code quickly, and conversely efficient code more slowly, so from a business value perspective it there is another factor which is the cost of running your process versus the value of running your process. Sort of the 'business efficiency' of the result.
Consider a goods economy comparison of the assembly line versus the craftsman. An assembly line uses more people but produced goods faster, that was orthogonal to the quality of the good produced. So the variables are quantity of goods over time (this gives a cost of goods), the quality of the good (which has some influence on the retail price), and the ability to change what sort of goods you make (which deals with the 'fashion' aspect of goods).
So what is productivity? Is it goods produced per capita? Or goods produced per $-GDP? Or $-GDP per goods produced? Its a bit of all three. Programmer productivity is just as intermixed.
That's not only false, but is often the opposite.
The symptom number one of an inexperienced programmer is to waste development hours reinventing the (square) wheel, while a good programmer is lazy (already knows which solution works best, and will probably just import it from a tested library).
So an experienced programmer not only doesn't waste computation power, also doesn't waste hours on the development cycle.
I agree with everything else you pointed.
My statement about inefficient code quickly is in terms of joules per computation. So while it is absolutely true that a junior perl programmer might slowly generate inefficient code and an experienced (lazy) perl programmer might quickly generate optimal perl code, neither of them would produce the same product written in assembly code (or better yet pure machine code).
To put that in a different perspective, I once wrote a BASIC interpreter in Java (one of my columns for JavaWorld) and it was pretty quick to do, and yet looking at the "source" to Microsoft BASIC written in 8080 assembler it was not very efficient. But it took Bill a lot longer to write Microsoft BASIC in assembler, and you couldn't even begin to port a full up Java VM to the 8080 (let's not argue about J2ME).
But step back then from that precipice, you have two versions of BASIC, one runs in a Browser and one runs on a 16 line by 64 character TVText S-100 card. (or 24 x 80 CRT terminal). Now you can run the same program in both contexts, unchanged, but the amount of energy you expend to do so varies a lot. So which is more "efficient?" I'd argue the one written in 8080 assembly is more efficient from a joules per kilo-core-second standpoint. Which was written more quickly? Mine, it only took about a week.
That is why talking about efficiency and productivity without getting anally crisp in your definitions can lead to two opposite interpretations of exactly the same statement.
[1] I find the lack of a wrist pad to rest on a challenge.
And to be fair we do - the great works of art, craft a d even science are almost always elegant. And it is something we atrive for. Some buggers have it and the rest of us wade around with feet of clay - but it is conceivably measurable.
Going to be easier to go with the editor's taste I think though.
It's a lovely idea, but in practice would be hard. "I just updated my interpreter and my code got 20% more efficient!" "How does it compare to that other code?" "I don't know, it's incompatible with the new interpreter."
While I could be accused of excessive cynicism, I do believe this is common enough that it should be addressed. There's a pervasive delusion that decisions are made by rational, informed actors, when that is rarely the case.
If the stakeholder you choose is a customer, then that is a valid measure of business productivity.
Which I guess is kind of the point - we are trying to measure on a granularity beyond what we can validly do.
Which indicates to me that a world of smaller organisations, made up of software literate people will be one where rewards will follow talent. That may not be a world we want to live in - and my cynicism sees your cynicism and raises :-)
1) Are you doing the right things?
2) Are you doing things right?
They can be imprecisely measured, but every metric has problems and can be gamed. Combining the measurements is extremely difficult.
Let's start with 1 - doing the right things. Someone who chooses to have their team work on 3 high value tasks, and stops their early on 6 low value tasks is by one definition more productive than someone who forces their team to do all 9 things. Or at the very least they are more effective. This is what Fowler is getting at.
On point 2... Let's assume that the appropriateness of what you are doing is immaterial. How fast are you doing it? This can be somewhat approximated. You can say "Speed versus function points" or "Speed versus budget" or "Speed versus other teams achieving the same output" and then bake in rework into the speed. All of these metrics are doable. Lines of code isn't a good base though.
The real question is, "What are you going to do with all of this productivity data?" If the answer is systemic improvement, you're on the right track. If you try to turn it into personal performance (or salary) then people wind up gaming the metrics.
Everybody says there's a "shortage of developers," but I know good developers who keep getting shitcanned after a few interviews where nothing seemingly went wrong.
We can't tell who's going to be productive. Since we can't tell, we come up with ten foot high marble walls to scale. Our sterile interview problems make us feel "well, at least the candidate can do our Arbitrary Task, and since we decided what Arbitrary Task would be, they must be good, because they did what we wanted them to do."
Productivity is pretty much the same. There's "just get it done" versus "solving the entire class of problems." Is it being productive if you do 50 copies of "just get it done" when it's really one case of a general problem? I'm sure doing 50 copies of nearly the same thing make you look very busy and generates great results, but solving the general problem could take 1/20th the time, but leave you sitting less fully utilized after (see: automating yourself out of a job).
The question I have is, how is this much different than any other profession? How do we measure doctor productivity? What keeps me up at night is that it is very likely that the 90/10 crap to good ratio in software developers is probably the same ratio as surgeons.
I am wondering if the ratio holds for crap to good parents. The scarier aspect of this is that people are actually being trained for their professions, as opposed to parenting, so the ratio may be even worse.
You know how I figure out if something can be improved? I dig in, understand it, and then look for ways to improve it. If I don't find anything, of course it doesn't mean there is no room, but I'm a pretty bright guy and my results are about as good as any other bright guy/woman.
I was subjected to endless amounts of this because I did military work for 17 years. You'd have some really tiny project (6 months, 2-3 developers), and they'd impose just a huge infrastructure of 'oversight'. By which I mean bean counters, rule followers, and the like - unthinking automatons trying to use rules, automatic tools, and the like. Anything to produce a simple, single number. It was all so senseless. I know that can sound like sour grapes, but every time I was in control of schedule and budget I came in on time and on to under budget. But that is because I took it day by day, looked at and understood where we were and where we needed to go, and adjusted accordingly. Others would push buttons on CASE tools and spend most of their time explaining why they were behind and over budget.
I like Fowler's conclusion - we have to admit our ignorance. It is okay to say "I don't know". Yet some people insist that you have to give an answer, even if it is trivially provable that the answer must be wrong.
If you're referring to Fred Brooks, he wrote "[T]here is no single development, in either technology or management technique, which by itself promises even one order of magnitude improvement within a decade in productivity, in reliability, in simplicity." (emphasis mine)
The surrounding context makes his comment a very specific prediction which means something different from what most people claim he meant. Much of the rest of his essay suggests techniques which address the issue of essential complexity and which, when applied together, he hoped would produce that order of magnitude productivity.
Perhaps there was no single such improvement in the years 1986 to 1996, but when people use the phrase "no silver bullet" to dismiss potential improvements in productivity, I believe they're doing Brooks and the rest of us a great disservice.
After all, much of programming culture is stuck on the idea that the clarity of syntax of a programming languages to novices is more important to maintainability of programs written in that language than domain knowledge, for example.
Now the more productive / better group is which can do the task with smaller complexity.
Complexity measures measure size of code and number of dependencies between blocks in different ways. But even the most simple comlexity measure is quite good: just measure number of tokens in source code. (It is a bitmore sophisticated than LOC). You can then make competitons between groups, and measure their productivity. (I am writing a book now titled 'Structure of Software' which discusses what is good software structure on a very generic/abstract level. It relates to 'Design Patterns' as abstract algebra relates to algebra.)
On the other hand, setting the time constraint (as opposed to measuring both time taken and solution complexity for the two groups) is important because deadlines help.
Chapter 8 "Beyond lines of Code: Do we need more complexity metrics?" by Israel Herraiz and Ahmed E Hassan.
Their short answer is that, in the case they looked at, all the suggested metrics correlated with LOC, so you may as well use LOC as it's so easy to measure.
IIRC they believe it's only good to compare LOC between different employees if they are doing pretty much the exact same task however, but since LOC is correlated with code complexity, there is some measure there.
I recommend the book, as really focusing on the science of computer science.
Measure it. Or optimize it. Can't do both without impacting the other.
Software is a work of art and creativity, not the work of a rules-based factory.
Basically I see this as marketing. We may not be the fastest but who cares about that we have the special insight to build the hits that keep you in business.
The problem with productivity measures, is not how they are measured but what they are used for. Most managers want to use productivity measures to evaluate individual or team performance, however, performance is tied to incentives, so you always end up with a lot of push back from the team or someone gaming the system. (IMO, this is because of lazy managers wanting to "manage by numbers", without really understanding how to manage by numbers.)
Rather than using it as a performance management tool, productivity measures, however imprecise, can be used alongside other yardsticks as signals of potential issues. For example, if productivity measure is dropping with a particular module/subsystem, and defect rate is increasing, then one might want to find out if the code needs to be rearchitected or refactored. In these cases, it is okay to be imprecise, because the data are pointers not the end goal. When used correctly, even imprecise data can be very useful.
In my opinion, we should approach measurement from a different angle: can we learn something useful about our profession by combining different types of measurements. Can we, for example, easily spot a person who is doing what Fowler is calling important supportive work. Can we detect problem categories that easily lead to buggy code and allocate more time for code quality work for tasks and less for those that are known to be more straight-forward.
You end up with something like feature 1: +12,544 / -237 lines. Done in 2 weeks.
Then comes feature 2, 2 and a half months later, the stats: +5,428 / -9,845.
Look at that, you had to tear down everything they wrote because they cared about amount of code over code quality. The more they brag, the more you think "oh s$%t, every line they add is a line I'm going to have to completely untangle and refactor."
I think software engineering productivity can be measured, though not well by today's standards. There will probably be a decent algorithm to do it in the future that takes in to account the power of the code, how easy it is to build on top of, how robust it is, etc.
How could that be possible? Consider the following, very typical scenerio.
You write 2000 lines of code, implementing a feature. I write 2,200 lines of code, implementing a feature in a way that supports our vertical for the next 5 years (you just take my module and plug it in, instead of coding from scratch). Add in whatever time interval you want to make it more complicated - I took the same amount of time as you, less, or more.
Consider that this is a judgement call - did I or you do the right thing? We understand the risks and costs of premature design, but also understand the risk of coding exactly to today's requirements, with no insight into the future. No algorithm is going to tell you the right answer, and by the time we know (5 years from now) the measurement will be useless.
Or we each write a heuristic to the TSP. What algorithm could possible decide whose work is more "productive"? I put it in scare quotes because I don't even know how to define productivity in that regard. Yours runs faster, mine took 1/2 the time to code. Yours is 70% larger than mine, which has cache coherence implications as we continue to add to our programs. Yours is well documented (give me an algorithm to tell if code is 'well' documented), mine is sparse. I used gotos, you used exceptions to deal with errors. You wrote it in Haskell, I did it in C++. Yours will make features X,Y,and Z easily possible, mine makes A,B, and C easy. Your heuristic performs better for some graphs, mine performs better on others.
Who is more 'productive' here? It's not even a meaningful question. Bottom line is, we both tackled a hard problem, both did fine in very different ways, and both have implications on the future of the company (assume a,b,c,x,y,z are really important in the future).
It's an N-dimensional optimization problem with endless unknowns, and no knowledge or agreement on how to measure many of the axis', let alone their relative importance to each other.
Now, I do like to look at my personal lines of code because it gives me a gauge to compare features I implement on a relative basis. It also gives me a relative, rough measure how much effort a particular feature took to produce.
http://folklore.org/StoryView.py?project=Macintosh&story=Neg...
True, it's hard to objectively measure the overall productivity using a universal standard, but it is relatively easier to measure the productivity fluctuation caused by the external factors. Velocity measurement in Agile practice is mostly for that end.
For the internal factors, the best way, and arguably the only effective way, to manage it is probably to hire good motivated developers. I think most top level software companies have learned that.
Add to this I don't think scrum has become setup to take this to its logical conclusions - agile/scrum has been sold as a fairly fixed methodology, not as a means to get some relative metric out of teams and use that in a series of experiments to achieve productivity improvements. And even if it were, the major wins we know and can prove work (quiet conditions, minimal interruptions, trust, respect, time for reflection and education, are a long way from being accepted by today's enterprises.
In short there is no silver bullet, and while agile looked a magic bullet it just turned out to be plain old lead.
> Copy and paste programming leads to high LOC counts and poor design because it breeds duplication.
This problem is not insurmountable. Compression tools work by finding duplication and representing copies as (more concise) references to the original.* The size of the compressed version is an estimate of the real information content of the original, with copies counted at a significantly discounted rate. The compressed size of code could be a more robust measure of the work that went into it.
* Sometimes this is done explicitly, other times it's implicit
Even if you deliver a system with a lot of features and no known bugs, if they aren't the right features, it's not valuable software.
http://en.wikipedia.org/wiki/Coastline_paradox
productivity of working on a software is like measuring fractals.
Count lines of code, function points, bugfixes, commits, or any other metric, and you're capturing a part of the process, but you're also creating a strong incentive to game the metric (a well-known characteristic of assessment systems), and you're still missing the key point.
Jacob Nielsen slashed through the Gordon's knot of usability testing a couple of decades back by focusing on a single, simple metric: does a change in design help users accomplish a task faster, and/or more accurately? You now have a metric which can be used independently of the usability domain (it can apply to mall signage or kitchen appliances as readily as desktop software, Web pages, or a tablet app).
Ultimately, software does something. It might sell stuff (measure sales), it might provide entertainment, though in most cases that boils down to selling stuff. It might help design something, or model a problem, or create art. In many cases you can still reduce this to "sell something", in which case, if you're a business, or part of one, you've probably got a metric you can use.
For systems which don't result in a sales transaction directly or indirectly, "usability" probably approaches the metric you want: does a change accomplish a task faster and/or with more accuracy? Does it achieve an objectively better or preferable (double-blind tested) result?
The problem is that there are relatively few changes which can be tested conclusively or independently. And there are what Dennis Meadows calls "easy" and "hard" problems.
Easy problems offer choices in which a change is monotonic across time. Given alternatives A and B, if choice A is better than B at time t, it will be better at time t+n, for any n. You can rapidly determine which of the two alternatives you should choose.
Hard problems provide options which aren't monotonic. A may give us the best long-term results, but if it compares unfavorably initially, this isn't apparent. In a hard problem, A compares unfavorably at some time t, but is better than B at some time t+n, and continues to be better for all larger values of t.
Most new business ventures are hard problems: you're going to be worse off for some period of time before the venture takes off ... assuming it does. Similarly, the choice over whether or not to go to college (and incur both debt and foregone income), to to learn a skill, to exercise and eat healthy.
It's a bit of a marshmallow experiment.
And of course, there's a risk element which should also be factored in: in hard problems, A might be the better choice only some of the time.
All of which does a real number in trying to assess productivity and employee ranking.
Time to re-read Zen and the Art of Motorcycle Maintenance.