July 2021
Monthly Archive
Monthly Archive
Posted by Jim DeLaHunt on 31 Jul 2021 | Tagged as: robobait, software engineering
I have been writing software for a long time. But every now and then I get flummoxed when my code misbehaves. It can take me a long time to figure out what is wrong. Often, it turns out to be a combination of basic mistakes which I have made, which remind me of some basic lessons in software engineering, which I didn’t realise that I was not applying. This happened to me recently. It is a story I call, “the exception”. Maybe, if I tell you the story, you can learn from my mistakes, and I can reinforce the lesson to myself.
I had been working on a set of command-line tools for a consulting client in recent years. The purpose of the tools is to take some input files, modify them in related ways, and write out the modified files. (This whole explanation is very simplified, to protect the client’s confidentiality, and to clarify the story.) Let’s say that I started off with a tool that would merge input file A and input file B, and write output file AB. Then as an experiment, I made a tool which would modify the contents of file B during the merge with A, and write output file ABʹ. I put this experiment away in a branch for quite a while. In the meantime, I extended the tool program so that it would take a directory D in place of file B, and merge file A with every file in D. So if D contained files D1, D2, D3, …, the tool merged and wrote AD1, AD2, AD3, etc.
The tools were a set of Python packages and modules. There was a module, merge_command.py, which was a command-line tool. There was also a package, modifiers, with four modules: common.py, to do the basic work common to all tools; and merge.py, to do the specific work of the merge tool. There were also other modules for other modifications, represented by shorten.py. They don’t matter for this story. Finally, like every Python package, modifiers had an __init__.py module to provide an interface that hides the internal structure of the package.
The code structure was something like this:
src/ merge_command.py # command-line program modifiers/ # package to do file modifications __init__.py # package interface common.py # basic work for all tools merge.py # specific work of merge tool shorten.py # specific work another tool
There were some further details which matter to the story. Each of the modules in modifiers defined a class to do its work: Common to do the work for all tools, Merge to do the work of the merge tool, etc. Class Merge and other classes inherited from class Common. So, merge_command.py gathered and checked command-line arguments, including paths for the input and output files, then passed them to an instance of class Merge to do the work. Merge in turn did some preparation to express the work in terms of what Common could do, then called superclass methods in Common to complete the work.
merge also defined some exceptions, represented here by exception InputFileNotFound. If class Merge could not find one of the files given in its path arguments, it raised this exception. (Classes like Merge never interact with the user; their caller has that responsibility.) merge_command.py caught this specific exception, and printed out an appropriate error message, and continued with any further work it could do. merge_command.py also caught any other exception. It would be unexpected and the tool had to stop. merge_command.py printed a general error message and stopped.
Finally, there was a test module for almost every code module. The only one that matters for this story was merge_command_test.py, which had tests of running the entire command-line tool. One test was test_missing_file(), which ran the tool with a path to an intentionally-missing input file, and checked that the tool printed the correct error message. I didn’t notice at the time, but this test contained my first error, which I will discuss below.
Thus, in more detail, the structure of the code was like this:
src/ merge_command.py # command-line program imports InputFileNotFound from modifiers imports Merge from modifiers creates instance of class Merge calls Merge instance with file paths etc. handles InputFileNotFound, prints specific message handles any other exception, prints general message merge_command_test.py # tests merge_command overall test_missing_files() calls merge_command with path known to not exist confirms that merge_command raised exception checks error message against expected text modifiers/ # package to do file modifications __init__.py # package interface imports Common from common imports Merge from merge imports InputFileNotFound from merge # more, but details don't matter for this story common.py # basic work for all tools defines class Common merge.py # specific work of merge tool imports Common from modifiers defines class Merge, inherits from Common if file missing, raise InputFileNotFound defines InputFileNotFound shorten.py # specific work another tool # details don't matter for this story
At first, only class Merge worked with paths to directories. Other classes in package modifiers worked with paths to individual files. But then I wanted other classes to work with paths to directories also. At that point, I decided to abstract out the parts of Merge which worked with directories and iterated through the files, into a new class Directories, which inherited from Common. I changed Merge to inherit from Directories instead. merge_command.py did not need to change, because package modifiers protected it from the details of the package implementation. This is normal object-oriented design and normal refactoring.
I copied merge.py into a new module directories.py, and began deleting the parts that were about merging rather than about directory handling. I now defined InputFileNotFound in directories.py. I added new entries to __init__.py to import class Directories and InputFileNotFound. I changed merge.py so that class Merge inherited from class Directories instead of class Common.
But as I created directories.py, I made my second mistake: I failed to delete the existing definition of InputFileNotFound in merge.py, and I failed to delete the import of InputFileNotFound in __init__.py.
However, all my tests passed. Particularly, test_missing_file() passed. Thus I knew that the command merge_command.py still printed the same error message, when a path it was given pointed to a file which did not exist.
At this point, the code looked something like this:
src/ merge_command.py # command-line program imports InputFileNotFound from modifiers imports Merge from modifiers creates instance of class Merge calls Merge instance with file paths etc. catches InputFileNotFound, specific message catches any other exception, prints general message merge_command_test.py # tests merge_command overall test_missing_files() calls merge_command with path known to not exist confirms that merge_command raised exception checks error message against expected text modifiers/ # package to do file modifications __init__.py # package interface imports Common from common imports Directories from directories imports InputFileNotFound from directories imports Merge from merge imports InputFileNotFound from merge # mistake! # more, but details don't matter for this story common.py # basic work for all tools defines class Common directories.py # directory handling for all tools imports Common from modifiers defines class Directories, inherits from Common if file missing, raise InputFileNotFound defines InputFileNotFound merge.py # specific work of merge tool imports Directories from modifiers defines class Merge, inherits from Directories # no use of InputFileNotFound any more defines InputFileNotFound # mistake! shorten.py # specific work another tool # details don't matter for this story
The code stayed in this structure for quite a long time. I had made two mistakes, but I wasn’t aware of them. My tests passed. My client was satisfied.
But recently, my client asked for improvements to the tool. These involved bringing the old experiment, wrote output file ABʹ, up to date with the current code. Specifically, I merged the current code of the main branch into the old code of the experimental branch. A number of modules conflicted, as I had expected. Many of the conflicts had to do with directory handling, and class Directories. The experimental code predated this new class. I worked to resolve the conflicts. This involved looking at the main branch’s code, and the experimental code, and their differences, and deciding what to take from the main branch, what to take from the experimental branch, and what to rewrite to reconcile the two. I wanted to have one set of code on the experimental branch which preserved all the current behaviour of the tool, and also had the experimental behaviour.
As I was resolving these conflicts, I noticed that __init__.py imported InputFileNotFound twice. The second import, from merge, was no longer correct. I deleted it. I then discovered the left-over definition of InputFileNotFound in merge.py, and deleted it also.
All of a sudden, the test, test_missing_file(), began failing. The output from merge_command.py was different than my reference expected text. I checked, carefully, for the edit when the problem arose. It arose when I merged the main code into the experimental. Perhaps I had made a mistake in merging the conflicts? But none of code for directories or missing files was in conflict, because it hadn’t existed in the experimental code. I ran merge_command.py in a debugger. The InputFileNotFound was raised, and handled, (apparently) as expected.
This kind of problem is confusing, frustrating, and demoralising for a developer. There is a temptation to run the same tests the same way several times, hoping for a different result. But that is insanity.
The way out of a problem like this is similar to the way out of a frozen lake after you have fallen through the ice: crawl gradually, remaining at a low level, testing assumptions, until on solid ground.
I read the exception handling in merge_command.py, and it was simple and very clear: there was code to handle an InputFileNotFound, and I could not see how it would fail. But then I used a debugger to step through the test_missing_file(), and was astonished. The Directories class raised an InputFileNotFound exception, as expected. The handler which should have handled the InputFileNotFound unexpectedly did not handle it. The general handler, to catch any other exception, handled it instead. When I examined the exception’s value, it looked just like an InputFileNotFound. But when I used the Python expression to see if it (ex
) was what it thought it was — isinstance(ex, InputFileNotFound)
— the answer was False
.
How could an InputFileNotFound exception not be an InputFileNotFound?
Eventually, the answer came to me. The name InputFileNotFound was defined twice: in directories.py, and an old left-over definition in merge.py. The Directories class used the name in directories.py to create and raise the exception. The modifiers package interface in __init__.py imported the name twice, the first time from directories, and the second time from merge. The second definition was what merge_command.py got when it imported the name from modifiers.
The InputFileNotFound handler in merge_command.py was matching against the exception defined in merge.py, rather than the correct one defined in directories.py. Even though they had the same name, they referred to different classes. The Computer Science name for this is “aliasing”. When used correctly, it is wonderful. When used unintentionally, it has effects that are really, really confusing.
My first software engineering mistake was to not delete the old definitions in, and imports from, merge.py when I extracted the exception to directories.py. If I had deleted the old definition, then the import in __init__.py would have caused an error, and I would have known to remove it. When I finally did this cleanup, as part of the branch merge, then merge_command.py finally got the correct definition for its InputFileNotFound name.
But that means that merge_command.py had been handling exceptions wrong the whole time. How had the tests passed? Well, that was my second software engineering mistake. I implemented merge_command.py first, then wrote the test_missing_file() code to check that it always handled the exception the same way it did at first. I failed to check the nature of that handling for correctness. I failed to check for the InputFileNotFound handler’s “specific message”. Instead, I took the (incorrect) “general message” as the reference correct output. This is sloppy test writing.
Thus, the failing test test_missing_file() was the sign of a change in the code under test, all right, but it was a change from broken to working!
So, I learned about software engineering from that. My three lessons are: