Premature Generalization

I’m sure most developers reading this post will be familiar with the term “premature optimization”. For those who may not be familiar, this refers to the tendency that some developers have to worrying about the performance of their code, usually but not always in terms of execution time, before their code is proven to work correctly in terms of customer requirements. It’s like saying, “Hmm, this function call may take 3 clock cycles to execute but if I inline the code it won’t take any extra clock cycles” before the developer has correctly defined the function to be called. It’s generally considered a poor way to write software. I believe it was probably Joe Armstrong, one of the co-creators of Erlang, who said “Make it right, make it fast, then make it pretty.” Getting the code right always comes before any sort of optimization should be attempted.

Risking the perils that may befall anyone who coins a neologism, I propose another premature condition which developers should try to avoid: premature generalization. I would define premature generalization as being the tendency to assume that what you see or what you want is what every other developer sees and/or wants. It’s the tendency of the novice developer to post messages like “Hey function x isn’t working. What’s wrong?” assuming that the rest of us also have the problem they have and we therefore don’t need any more detail to know what he or she is talking about. It’s also the tendency to assume that because they need to address some particular use case that most developers working in the language will need to address the same use case: “Hey I need library function x to return an error status when I pass it a bad argument but throw an exception when I pass it a good but out of range argument. Can we modify the standard library to do this?” In both cases, I believe they’re prematurely assuming their question is generic and therefore they don’t need to provide details to others.

Although I feel I’m flogging a dead equine I hasten to point out that as software developers we should usually proceed from the assumption that the error we’re seeing is due to something specific to us. We should also proceed from the assumption that any use case we have is likely to be a use case specific to us until we see evidence otherwise. Yes, I am aware of the irony of making generic statements here—but given what I’ve seen over 25+ years of software development I don’t believe that I’m indulging myself in premature generalization.

Quick Elixir Debugging Tip

HT: Dmitri Skliarov for sharing this tip with me.

You can quickly see the call stack being invoked when you call a given Elixir function via a quick and simple call out to the Erlang DBG module. The following tip is a very simple example of how to use this; there’s quite a bit more that can be done with DBG but you need to look at the module docs to see them.

First a little code that we want to trace:


defmodule Trace do

def add(n, m), do: n + m

end

Next you will want to initialize the debugger:


:dbg.tracer()

# => {:ok, #PID<0.66.0>}

NB: you are extremely unlikely to get the same PID I’ve shown here. Next we need to tell the debugger which item or items we want to trace:


:dbg.p(:all,:c)

# => {:ok, [{:matched, :nonode@nohost, 45}]}

Next we tell it precisely what we want to track:


:dbg.tpl(Trace,:add,:x)

# => {:ok, [{:matched, :nonode@nohost, 1}, {:saved, :x}]}

NB: the first parameter to tpl is actually an atom but since a Module name is automatically an atom it’s fine to pass the module name without a colon in front of it. Also note that you can pass the special atom :_ as the second parameter to trace all calls to all functions.

And finally we actually trace our code:


Trace.add(1,1)

# => (<0.57.0>) call 'Elixir.Trace':add(1,1)
# => (<0.57.0>) returned from 'Elixir.Trace':add/2 -> 2
# => 2

And for reference when we use :dbg.tpl(Trace,:_,:x):


Trace.add(1,1)
(<0.57.0>) call 'Elixir.Trace':'__info__'(macros)
(<0.57.0>) returned from 'Elixir.Trace':'__info__'/1 -> []
(<0.57.0>) call 'Elixir.Trace':add(1,1)
(<0.57.0>) returned from 'Elixir.Trace':add/2 -> 2
2

 

And finally when you’re done tracing:


:dbg.stop_clear()

There’s a lot of power using this technique so I’d advise that you take some time to look at the DBG man page before you start using this extensively.  Also, it’s my understanding that you should always be careful to use this technique only in development because if you trace the wrong thing you can cause deadlocks.

 

 

 

 

Quick Elixir Tip

I use Samuel Tonini’s Alchemist (and Elixir mode) to make myself much more productive with Elixir and Emacs.  But one thing that I’ve found to be a minor annoyance is that whenever I start up Emacs I have to set the scratch buffer to be in Elixir mode and I then need to change the comment characters.

Thanks to this excellent tip I no longer have to do this.  I’ve modified my ~/.emacs to add these lines to the bottom:


(setq initial-major-mode 'elixir-mode)

(setq initial-scratch-message "\
# This buffer is for notes you don't want to save, and for Elixir code.
# If you want to create a file, visit that file with C-x C-f,
# then enter the text in that file's own buffer.")

And when I get a scratch buffer I’m all good to go with Elixir!

Making Exrm Work On Windows

So as many of you will know I’m very interested in Elixir.  I also want to see it work better on Windows.  There are a few of us that have been working on it.

I’ve also been very interested in how one can deploy an Elixir app on a machine which doesn’t have Erlang and Elixir already installed.  While most of us don’t mind having to have Erlang and Elixir around because we are doing development work, it would be much more secure to only have to deploy the binaries that need to be deployed for an app.

Since I started on Elixir one of the nicest folks in the community (and that’s saying something) has been Paul Schoenfelder. When I was first struggling to learn Elixir, Paul very patiently helped me to get started.  So I became aware of one of the utilities that Paul built: EXRM.  Exrm is a tool to automate the production of a package of the needed Elixir and Erlang binaries to deploy an app. It allows a developer to deploy his or her Elixir app with only the needed runtime files so that the target machine doesn’t have to have Elixir or Erlang installed.

So I asked Paul what I could do to sort of return the favor for the help he gave me when I started and he mentioned that he wasn’t as pleased as he might be with the support for Windows in EXRM.  So I started digging into EXRM to see what I might do to get it working on Windows.

A few instructive things I spotted right away. I noticed that the way the Windows batch files were constructed they will not work correctly if they’re not run from the correct directory. I mention this for others who may try to use EXRM on Windows.

Spaces In Directory Names

I noted also that there were several places where directories containing spaces and too many characters were also a problem. The original developers of the batch files wrapped double quotes around all the directory names. If you decide to build Windows batch files, please note that wrapping double quotes around directory names is a bit error prone and tough to get right.  Hence I opted to modify all the various places where they had double quoted directory names to using 8.3 versions of the names.  That fixed up all the references to directories so that everything was found correctly.

Getting Services Working

Next there was the question of getting Windows Services working for Erlang. While digging into the batch files I discovered that basically the issue hinged on the installation of the service.  Once the service was installed correctly, it looked as if the rest of the operations on the services (starting, stopping, restarting etc.) would work correctly.

I had to hunt around for a while to figure out what I needed to do to get the service to install correctly.  I finally figured out that what was happening was the wrong parameter was getting passed as the binary to be started when the service gets started.  And so I repaired it.

A Few More Details

You’ll also need to unblock the erl.exe and epmd.exe files for the Windows Firewall.  You can wait for Windows to prompt you or you can proactively unblock them via this escript.

You may also need to deploy msvcr100.dll to your deployment machine as well.  If it’s a Windows 10 machine you may need to do this.

So we’re almost ready to roll the changes into EXRM so we can get people to be able to deploy Elixir apps on Windows.

I will continue to work with Paul on improving support for EXRM on Windows.  Keep watching this space; we’ve got some interesting ideas we’ll be trying to roll out in the next few months or so.

 

 

 

Three Rules Of All Testing

There are three rules applicable to all testing environments.  I’ve articulated one of the rules before but now I’ll add two more.

Rule 1 – Dead Simple

Tests must be dead simple to run.  It doesn’t matter what you’re testing.  It doesn’t matter if you’ve got 100% code coverage. It doesn’t matter if you’ve tested everything from the smallest unit to the largest UI.  If the tests aren’t easy to run the people who most need to run them that is the developers and the analysts will figure out ways to game the system and avoid running the tests.  The rule is that running tests must be easier than any workaround the developers and analysts can come up with to avoid running the tests.

Rule 2 – Some Tests Are Better Than No Tests

I can’t count the number of times we’ve been discussing testing and someone will say “well, we can’t test X” with the implication being that since we can’t test everything there’s nothing to be gained from any testing.  This is analogous to saying the only way we can engineer a rocket is to build the whole thing and then fire it off.  Never mind building prototypes or doing any of that silly math to check our assumptions.

Now, granted there can be some cases where certain properties are so essential to a system that if they can’t be tested it really isn’t worth testing at all.  But those are extremely rare cases.  Even if you can only run tests on 50% of your code, that’s still 50% that’s tested that would not get tested otherwise.

Rule 3 – Don’t Test the Language or the Libraries

I’ve seen unit tests where someone will set some property of a class.  Then they will immediately read the property to check that is hasn’t changed.  While it’s a valid test to check an invariant on a class, it’s not valid to test it immediately after its set.  If you set a value and then read it and it’s different, you’ve got problems more substantial than a unit test will ever solve.

These are just my big three rules for testing.  I’d be interested to hear the thoughts of others on this subject.

Follow

Get every new post delivered to your Inbox.

Join 629 other followers