Effectiveness

Various notes to myself, about being effective.

1.  allow for some slop in the machinery.

Being intolerant of error or imprecision in others (or in one's self) lowers effectiveness.  To err is human.

2.  strive for rightness, but don't beat people over the head with it

Related - being right doesn't always mean you're effective.  And being right in a way that pisses off people around you (e.g. right in a pedantic sense) lowers your effectiveness when people disengage from you.  Often as geeks we assume that our individual contribution is really all that matters; but on most projects, our contribution as a member of a larger hive is as important (maybe more so, on some projects).

3.  furious activity is not a substitute for genuine progress.

Charging full speed down the wrong road, doesn't lead to being effective.  The cure (if there is one) is to understand the project's goals well. 

4.  but at some point you need to stop staring at your navel.

The flip side of [3], is that you also must to realize when you need to actually deliver something (as opposed to spending more time in "understanding the problem" better).

5.  consistently avoiding risk is a net loss.
 
If you're in a position where covering your ass (reducing project risks that you could be blamed for) is more important than delivering running code on time, it's probably a good time to find a different project to work on.

6.  listen to customers.

It's a good thing to listen to customers.  To listen, of course, implies you have a relationship with customers in the first place.  "I don't do customer facing stuff" is a good way to tell everybody "I'm obsolete."

7.  ideas are good.  implementable ideas are vastly better.


No one will pay me to come up with ideas that I have little or no chance of implementing.  Work out a rough sense of what can actually be built, or demo'd, and find customers who would be willing to build on that.

The Classic Blunders

I was thinking today about The Classic Blunders, as described by Fezzini.

1.  Turnover isn't free.  So much knowledge never gets written down; it's domain knowledge or experience that only exists in the mind of the person.  When the person walks out the door, all that know-how goes with him.   (And that's just the smallest cost.  Pour encourager les autres is a great way to utterly demoralize les autres.)

2.  Prickliness isn't free.  When you have prickly people you have to deal with, day in and day out, it wears on the soul.

3.  The Mission Isn't Everything.  When The Mission Becomes Everything, it means people are willing to sacrifice all the essential parts of a reasonably happy and healthy life - like health, family, a sense of balance - In Order To Make The Mission Happen.

A tale of two distances (part 2)

Last time,  I wrote about using cosine similarity as a way to tell how alike or different two trigram language models were.   This post describes a second distance-like measure, from Cavnar and Trenkle's 1994 paper, "N-Gram-Based Text Categorization", which can be found on the web.  The way Cavnar and Trenkle (C&T) compute the distance between models is not via cosine similarity, but by using the ranks of the sorted trigram frequencies.

Here's a simple example of C&T's method:  suppose you have two lists of trigrams, and we want to make a distance measure that will tell us how alike the lists are.

list X:    RAN CAT CAT THE THE THE RED
list Y:    THE CAT RAN RED RAN CAT RED


We compute the frequencies of the trigrams in the lists, sort the lists in descending frequency, rank the resulting lists, and compute the sum of the differences of the ranks.  As worked out:

X freqs            RAN=1 CAT=2 THE=3 RED=1
sort down by freq  THE   CAT   RAN   RED
rank               1     2     3     4

Y freqs            THE=1 CAT=2 RAN=2 RED=2
sort down by freq  CAT   RAN   RED   THE
rank               1     2     3     4

rank diffs         THE(1-4) CAT(2-1) RAN(3-2) RED(4-3)
absolute diffs     3        1        1        1
sum of abs diffs   6 = 3+1+1+1


Here, lists X and Y are trigrams.  We compute the differences between the ranks of the lists as sorted in descending order by frequence, then take the absolute values of these differences, and sum these absolute values.   The resulting sum (6 in this example) is a measure of the mismatch between the two lists.    

Cavnar and Trenkle call this measure of mismatch an "out-of-place measure".  In this posting, I call the measure of mismatch the rank distance between the two lists.
When an element is in both lists, the magnitude of the rank differences for that element tells us how far out of place that element is.   When an element is not in both lists, we arbitrarily set the rank  difference to N (the length of the lists).   The values of rank distance vary from 0 (both lists have the same elements with the same frequency rankings) to N*N (the two lists share no common elements).

For two trigram models (from a reference text and a sample text), we can use the Cavnar & Trenkle out-of-place measure as a distance-like way to test the similarity between two models -- lower rank distance implies greater similarity.

The process described above is not exactly what Cavnar & Trenkle proposed.   C&T collect frequencies for all N-grams (N=1,2,3,4,5), not just trigrams and take the top 400 most-frequent N-grams in descending frequency order as their language models.  I've tested a simplified variant of C&T (only use trigrams, take the top 300 for a language model) against the Python Recipes code, and the quality of classification results are similar, when used just for language detection.

For reference, the C&T 1994 paper is available from

http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.162.2994

A tale of two distances

Suppose you have a blob of foreign text, and you want to identify what language the text is written in. At the Python Recipes web site, the URL

http://code.activestate.com/recipes/326576/

gives Python2 code for doing this.  The idea is simple.

For each language you want to detect, you build a reference model for that language:

1. get a reasonably large sample of text in this language.
2. from this sample, make a frequency table of the 3-grams in the text.
3. save this frequency table as a model for this language.

For a new sample text you want to classify:

1. make a frequency table of 3-grams (a new language model)
2. compute the distance between this model and those of all the reference models you already have.
3. classify this sample text as the language of the reference model with the least distance.

What do we mean by distance in the context of trigram language models?  

In high school, we were introduced to Euclidean distance - a way of measuring how far two points in the (x,y) plane are from each other:

distance(x,y) = sqrt((x1-y1)**2 + (x2-y2)**2) where x = (x1,x2) and y = (y1,y2)

So our garden-variety idea of distance is a function that takes two points (two ordered pairs), and returns a nonnegative number.   Math people call a function like this a metric (or distance function, or just a distance) if the function has the following properties:

1. coincidence.   d(x,y) = 0 if and only if x = y.
2. symmetry.  d(x,y) = d(y,x).
3. triangle inequality.   d(x,z) <= d(x,y) + d(y,z).

The Python Recipe code uses cosine similarity instead of a distance - if two language models are the same, they get a similarity value of 1.0, and if they are entirely unalike, they get a similarity of 0.  In the following description,A and B are frequency tables whose indexes are trigrams, and whose values are the frequencies of the trigrams in two text samples.  (Suppose A is one of the reference language models you built to test against, and B is the model from the sample text you want to classify.)

similarity(A, B) = 
    total = 0
    for each trigram t in both A and B:
        fa = A[t]    (frequency of trigram t in model A)
        fb = B[t]    (frequency of trigram t in model B)
        total += (fa * fb)
    return total / (scalar_length(A) * scalar_length(B))

and

scalar_length(X) = 
    return square root of sum of squares of all trigram frequencies in X.

The Python code implements roughly the pseudocode above; the trigram frequency tables in the Python are stored as dictionaries of 2-grams whose values are (dictionaries of 1-grams whose values are frequencies).

Rather than provide a distance, the Python code defines a difference between models, defined as

difference(A, B) = 1 - similarity(A,B)

This difference function acts vaguely like a distance, but actually is not one in the precise mathematical sense (see wikipedia article on "cosine similarity" for the reasons why).  In practice, though, the difference function acts the way we'd like a distance function to work:  two trigram-identical language samples yield a difference value of 0, and two entire dissimilar samples yield a difference value of 1.

Next time, I'll describe a different method for calculating a distance between two models, based on a nonparametric approach called rank distance.

When correctness is not enough

Met the spec, but not done yet

Unknown nonfunctional requirements are among the banes of a working programmer's existence.  ("Heck," I can hear you say, "lack of any requirements...")

I'm focusing on nonfunctional requirements because it's been more typical in my experience to have the customers give a few vague verbal clues which can be generously described as functional requirements, but the nonfunctional aspects of what the customer needed are just glossed over.    Functional requirements give us some warm fuzzy feelings as programmers, because we associated them with correctness, and the correctness of an artifact w.r.t. a spec is something we can actually establish (even if imperfectly).

The usual thing people mean when they say "this software module is correct" is "given the appropriate inputs, this module produces the expected outputs."   This everyday meaning corresponds directly with the functional requirements of the module.   Alas, software may have nonfunctional requirements as well.

When software meets the user's functional requirements but fails to meet the user's nonfunctional requirements, correctness alone is not enough. [+]   The software I build can produce the right results all of the time, and still be useless if it can't run in the customer's environment, or within the time constraints the customer hasn't told me about.

It's ok to say "well, that never happens to me."  Just don't be surprised if some of your customers are reluctant to continue paying you.

When people want software to solve a problem, I think there are a couple of quantum levels of capability that "plain old non-programmer people" can imagine:

1. I have zero software to help me do this thing I want to do.  If I can do this thing at all, I do it using a manual process.
2. I have some software to help me do this task, and the speed and resource consumption of the software doesn't matter.
   
3. I have software to help me do this task, and the speed and resource consumption of the software matters greatly, because I have real-world limits on how much time or resources I can throw at my problem.

The quantum leaps happen when a customer goes from one capability level up to the next one.   The moment when "correctness is not enough" is when you have correct software which is nonetheless useless, because the software doesn't meet the user's nonfunctional requirements - e.g. it runs too slowly or consumes too many resources.

The First Leap : Basic Automation

When you (non-programmer paying customer) go from Capability Level 1 to Level 2, this is generally a Big Deal.  A manual process (tedious, boring, error-prone) is now an automated process (still may be error-filled, but likely a lot less tedious and boring for the human user).  A lot of everyday business process automation deals with this jump from Level 1 to Level 2.   These are the sorts of problems where simply having a computer being able to do the task at all, represents a big savings in cost for the customer.

Ideally, your Level 2 software process correctly does what you need it to do.   Part of the immense work required in moving from Level 1 to Level 2, is that possibly for the first time, the actual Level 1 process has to be formally described - completely, clearly, and correctly.   This formalization may appear in part as requirement documents, but from the working programmers' seat, the real deal is the code that describes exactly what the computer is going to do.
But a lot of Level 2 software customers can live with some measure of incorrectness, as long as the presented functionality still brings a net gain in economic value.   (Clearly some customers want provable correctness guarantees, and are willing to pay for them.  I think these represent a minority of paying customers.)

The Second Leap : When Resources Matter

When you go from Level 2 (resources don't matter) to Level 3 (resources do matter), this is another quantum leap.    Before now, you (paying customer) had some software that did some useful task for you, but something has changed in the resources you need to run your software.

Some possible causes -  

• you acquired lots of new customers.
  
• the acceptable level of risk associated with your software has changed.
  
• the regulatory regime changed, which means you have to produce a lot of new metadata or audit trail info that you didn't used to produce.
• the competition changed, and now has something that competes with your original automated process, but offers a better economic value to your customers.
  

Whatever the cause, now resource consumption of your software matters in ways it didn't before.  (There are different degrees of "matters" - this is not an all or nothing proposition.)   Software to solve this problem now has to (a) be sufficiently correct, and (b) operate within the allocate time or resource limits.

Software in a correctness vs constraints space

One way to look at the relationship between correctness and resource constraints, is to picture software as existing as a point in a two-dimensional space:

In this graph, points A and C represent software with minimal resource constraints, and varying needs for correctness.  Points B and D represent software with strong resource constraints (e.g. speed or memory) and varying needs for correctness.   If you are point C, and you need to be at point D, then simply being correct will not make your software actually usable by the customer.   It's a truism that "performance doesn't matter, until it does", but for some customers, when performance matters, it really matters.

[+] To say "correctness is not enough" is not saying correctness doesn't matter at all -- of course it does.  Even for customers who use somewhat incorrect software (most of us, most of the time), at some point of perceived incorrectness, a paying customer who wants your software to solve his problem, stops paying for it, because the incorrectness of the proposed solution is too expensive to live with.

Stone Knives and Bear Skins

In "The City on the Edge of Forever", Spock explains to Edith Keeler that he's trying to make a mnemonic circuit using only stone knives and bear skins.

Today I'd like to show one way to build a simple make-like utility, using the Unix equivalent of stone knives and bear skins.

This is not a "code golf" exercise, in that I'm not trying for the smallest possible solution (however inscrutable). Instead, I want to illustrate how little machinery it takes to construct a tool that approximates what make(1) does.

Here are the ingredients:

1. a main shell script (make.sh), which needs
   • a modern Bourne shell, with shell functions
2. a set of shell functions, which need
   • cat
   • sum
   • awk
   • touch
   • /bin/test (aka "[")
   • a compiler (for today, I'll use 'gfortran' on linux)

and that's it. With some more cleverness, we could probably figure out how to use fewer unix tools than these, but this set of programs (sh, rm, cat, sum, awk, touch, and [ ) are commonly found on Unix hosts. We could substitute another compiler without much difficulty.

For concreteness, the target program I want to compile has two small files: main.f90 and print.f90. They are written in Fortran 90. Don't worry if you don't know Fortran -- the target code is just an example. You could substitute different code with the appropriate compiler, in a straightforward way.

main.f90 is the initial procedure of the program:

---cut here---
program main 
    implicit none
    integer, parameter :: limit = 5
    call print_table(limit)
end program
---cut here---

print.f90 contains the PRINT_TABLE subroutine:

---cut here---
subroutine print_table(limit)
    implicit none
    integer :: limit 
    integer :: i
    do i = 1, limit 
        write(*,*) i, i**2, sqrt(real(i))
    end do     
end subroutine
---cut here---

(A subroutine in Fortran is roughly like a function returning void in C.)

The finished program writes out a short table of squares and square roots. The first column is an integer, the second column is the square of the integer, and the third column is the square root of the integer.

$ ./main
           1           1   1.0000000    
           2           4   1.4142135    
           3           9   1.7320508    
           4          16   2.0000000    
           5          25   2.2360680    

Now to work. The aim is to write a shell script that acts like make(1) for compiling this program. Performing the compile and link steps one file at a time, our utility should do:

# compile source files into objects

$ gfortran -c -o main.o main.f90
$ gfortran -c -o print.o print.f90

# link objects into executable

$ gfortran -o main main.o print.o -lm 

If a source file changes, we want to recompile just that file. If any object file changes, we want to re-link the main executable.

For make(1) and its cousins, there are different ideas about what constitutes "change" in a file. Every useful make-like tool doesn't actually compute an absolutely true answer to the question "has file XYZ changed?" - instead, each tool computes an approximate answer.

In the case of make(1), a file is considered "changed" if the file's mtime has changed. That's an approximation that's fast and easy to compute: do a stat(2) on the file, and read the file's mtime. There are other approximations for "has file XYZ changed?" that you could use - the file's length, a checksum of its contents, etc. Each approximation has its pros and cons, and its fans and detractors.

For the tool I'm writing about today, I will use a checksum computed by sum(1).

The logic of the main build script "make.sh" looks like this:

        for each object file
                if this object or its source file have changed
                        recompile to make this object

        if any of the objects have changed
                re-link the objects to make the executable

Here's the source of the build script "make.sh" :

---cut here---
#!/bin/dash

export PATH=$PATH:. 

. ./funs   # source our helper shell functions

target=main

objs="main.o print.o"

for o in $objs ; do 
        dep=${o%.*}.f90
        if [ $(changed $o $dep) ] ; then 
                do_compile $o $dep 
                record_sum $o $dep
        fi
done

if [ $(changed $target $objs) ] ; then 
        do_link $target "$objs"
        record_sum $target "$objs"
fi
---cut here---

The remaining magic lies in the shell functions we called: 'changed', 'do_compile', 'do_link', and 'record_sum', which are all in a helper file of shell functions, here named "funs" :

---cut here---
# helper shell functions

checksum(){
        cat $* | sum | awk '{ print $1 }' 
}

record_sum(){
        checksum $* > $1.sum
}

changed(){
        local new old rv
        if [ -f $1 ] ; then
                new=$(checksum $*)
        else
                touch $1
                new=$(checksum $*)
        fi
        old=x
        [ -f $1.sum ] && old=$(cat $1.sum)
        rv=""
        [ $new != $old ] && rv=1
        echo $rv
}

do_compile(){
        if [ -f $1.sh ] ; then
                # compile using a specific build script
                $1.sh $*
        else
                # compile using generic logic
                echo gfortran -c -o $1 $2
                gfortran -c -o $1 $2
        fi
}

do_link(){
        local targ
        if [ -f $1.sh ] ; then
                # link using a specific build script
                $1.sh $*
        else
                # link using generic logic
                targ=$1
                shift
                echo gfortran -o $targ $*  -lm 
                gfortran -o $targ $*  -lm 
        fi
}
---cut here---

Let's look at each of these helper shell functions in turn, and work out what they're doing, and why they're doing it.

First, we have a shell function to compute a checksum of a set of files given as arguments:

checksum(){
        cat $* | sum | awk '{ print $1 }' 
}

This shell function contains all the magic needed to figure out if a file has (approximately) changed. You could substitute 'cksum' for 'sum', if you wanted to have a stronger approximation for "file changed"; or you could use 'md5sum' for an even stronger approximation.

Next, we want a way to record the computed checksums, to see what's changed from one invocation of "make.sh" to the next:

record_sum(){
        checksum $* > $1.sum
}

The 'changed' shell function is the real work-horse. A file is considered changed if either (a) the file doesn't exist, or (b) the file or any of its prerequisites' checksums have changed.

changed(){
        local new old rv
        if [ -f $1 ] ; then
                new=$(checksum $*)
        else
                touch $1
                new=$(checksum $*)
        fi
        old=x
        [ -f $1.sum ] && old=$(cat $1.sum)
        rv=""
        [ $new != $old ] && rv=1
        echo $rv
}

Finally, there are shell functions to do the actual compilation and linking:

do_compile(){
        if [ -f $1.sh ] ; then
                # compile using a specific build script
                $1.sh $*
        else
                # compile using generic logic
                echo gfortran -c -o $1 $2
                gfortran -c -o $1 $2
        fi
}

do_link(){
        local targ
        if [ -f $1.sh ] ; then
                # link using a specific build script
                $1.sh $*
        else
                # link using generic logic
                targ=$1
                shift
                echo gfortran -o $targ $*  -lm 
                gfortran -o $targ $*  -lm 
        fi
}

The logic in 'do_compile' and 'do_link' merit further explanation. In do_compile, we do one of two concrete things: either we compile the target using some generic logic, or we use a target-specific shell script (if this script exists) to compile the target in a special way.

Suppose we have a file called foo.f90. The corresponding object file is foo.o. The generic logic that turns a .f90 into a .o, looks like

gfortran -c -o foo.o foo.f90 

But if we wanted to override this generic logic, for just target foo.o -- i.e., we want to use different compiler options -- we make an executable shell script called 'foo.o.sh', that contains the specific command we want to use to compile foo.f90 into foo.o :

---example foo.o.sh---
#!/bin/dash
gfortran -O -c -o foo.o foo.f90 
---cut here---

The logic of 'do_link' is like that of 'do_compile' -- by default we do a default link action, but if a special link script exists, we run that to perform the link in a non-default way.

That's it. We have the main make script "make.sh", and a file of helper shell functions "funs". To make the executable, we run make.sh:

$ ./make.sh
gfortran -c -o main.o main.f90
gfortran -c -o print.o print.f90
gfortran -o main main.o print.o -lm

We could build additional scripts to implement other 'make targets', like a 'clean' target that cleans everything up:

---example clean.sh---
#!/bin/dash
rm *.o *.sum main 
---cut here---

Although this set of scripts (make.sh + funs + clean.sh) works for this exact program, you would have to rewrite parts of make.sh + clean.sh to work for other programs or languages that you use. That may sound like a big bother, but really these scripts have much stereotypical content that you can copy from project to project, and they're not that big (less than 30 lines total).

The trio of scripts (make.sh + funs + clean.sh) weighs in at just under 100 lines of code. Gnu Make, which is very featureful, has around 31 KSLOC.

The build tool we've made works in a top-down way, unlike make(1) which works in a bottom-up way. make(1) is "fed" with a declarative specification of what needs to happen, whereas "make.sh" has a very imperative way of doing the same basic task - building an executable from one or more source files.

Armed with this blog post, should you now abandon make(1) or your current build tool of choice? Probably not. Any build tool you currently use is likely to give you faster builds than the approach I've sketched above. That said, it's neat that a very small amount of simple shell scripting can do a useful subset of what make(1) can do - rebuild a software project from source, recompiling just the files needed.

I developed these scripts after reading about apenwarr's "redo" utility, and after reading Alan Grosskurth's master's thesis, "Purely top-down software rebuilding", which is referenced at apenwarr's redo site: https://github.com/apenwarr/redo .