On Information Loss in Software

“Information Loss” is a way to look at the world. The topic is very broad. This blog post will focus on information loss during development and operation of computer software.

This post discusses why Information Loss is bad and gives some examples.

My hope is that after reading this post, you will be able to spot information loss more easily. This should help you avoiding information loss, eliminating the need for costly information recovery phase. Some examples include specific recommendations how to avoid that particular case of information loss.

Information Loss Definition

Information Loss for the purposes of this blog is the situation where information I is available and is easily accessible at point in time t1 but later, when it’s needed at point in time t2, it is either not available or not easily accessible.

The post will present various categories of information loss with examples. The list is not exhaustive; it’s not meant to be. The intention is to give some examples to help you get the feel and start looking at things from the information loss perspective.

Why Information Loss is Bad?

In many cases of Information Loss, the missing information can be recovered but that requires resources to be thrown at the issue (time and/or money). That is the situation I would like to help you to avoid.

Between the Head and the Code

When working on software, the first place the information loss occurs is when the programmer translates thoughts into code. Information loss at this stage will manifest itself as increased WTF-per-minute during code review or just code reading. Each time the code is read, there will be additional cognitive load while the reader reconstructs the programmer’s idea behind the code.

I have identified two main causes for information loss at the head-to-code stage:

  • Programmer’s fault
  • Programming language imposed

Information Loss due to Programmer’s Fault

The more a programmer is experienced, the less likely is the occurrence of information loss at this stage.

Misnamed Variable

In programmers head: number of servers running the ETL task. Name of the variable in the code: n. WTFs at code review – guaranteed.

Misnamed Function

I’m pretty sure getUser() should not update say last name of the user in database. Such naming is criminal but unfortunately I’ve seen code similar to that.

Use of Magic Numbers

if (result == 126) .... The person who wrote 126 knew what that number means. The person reading the code will need to spend time checking what that number means. One should use constants or enums instead: if (result == NOT_EXECUTABLE) ....

Missing Comments in Code

Most important comments are about why something is being done as opposed to how. If ones code is in a high-level language and of a good quality, it’s a rare occasion one needs to comment about what or how something is being done. On the other hand comments like “Working around API bug: it returns false instead of empty array” are very valuable.

Incorrect Usage of Data Types

A list of people, for example, is not just a list. It has semantic meaning. It’s much easier to understand a program when correct types are used for the data. Java has generics to convey such information, for example List<Person>. Some other languages have type systems that are powerful enough to convey such information too.

Programming Language Imposed Information Loss

Limitations of programming languages lead to less expressive code because the idea in programmer’s head can not be expressed in a straightforward manner. The readers of the code will struggle more (read waste time) to understand the code.

Unnamed Function Parameters

bash and perl5 (not sure about perl5 anymore, there was something experimental) do not have the syntax for specifying function parameter names. This makes the code less expressive. Sometimes programmers will do “the right thing”:

myfunc() {
    local target_file=$1
    ...
}

… but when they don’t, you finish with unnamed parameter, wondering what it could mean:

myfunc() {
    if [[ -f $1 ]];then
        ...
    fi
}

Is that a file to generate or a source file? You don’t know, you have to read on in myfunc hoping for the answer.

Recommendation: even if your language does not support named parameters, emulate them.

Expansion of Strings into Several Arguments (bash)

rm $x

Does that remove one file or several? What the programmer meant? You simply don’t know. It depends on the contents of x, which is typically split into arguments by spaces. You are lucky if you can deduce from the variable name whether it’s one or several files.

From today’s perspective this is just bad design. Back at the day I guess it was the most practical way to implement arrays.

Recommendation: use one of the two alternatives blow and do not use rm $x form.

  • Single file: rm "$x" (proper quoting)
  • Multiple files: rm "${my_files[@]}" (bash arrays)

Side note: this “feature” caused so much pain over the years when x would contain a spaces by accident. Even when x is meant to be used as an array, elements of that array can also contain spaces by accident.

Error Handling

In languages that do not support exceptions (bash, C, Go), the programmer is forced into one of two situations:

  • Write incorrect code that ignores the errors (on purpose or by mistake, go figure which one)
  • Write verbose code that handles the errors. When the code handles every possible error, it becomes cluttered with error handling and it takes more time to understand the code. That’s the case where information loss occurs because the reader is overwhelmed by the code.

In NGS, since typical use case is scripting, I wanted to have the option for the code to be concise. That rules out returning status code along with the result because the caller is then forced to check it. It does make more sense for NGS to have exceptions and for scripts to decide whether to catch them or let the whole script terminate with error because of an uncaught exception.

Unordered Hash/Map/dict Data Structure

Hash data structure is implemented in a non-order-preserving manner in some languages. That means that the programmer can not express the intention freely in situations where the order of key/value pairs is important. That pushes towards less readable code as the programmer fights the language by implementing his/her own ordered dictionary.

Information loss in this case is again losing the sight of programmer’s intention.

Fortunately many modern languages solved the issue by now:

Recommendation: check whether your language has the data structure you really want to use, either built-in or in a library.

Limited Data Structures (bash)

Working with data structures in bash results more or less convoluted code, depending on the data structures one need to work with. This is direct consequence of bash supporting exactly three data structures:

  • Scalar (strings which can sometimes be treated as numbers or arrays)
  • Array
  • Associative array

These data structures can not be nested.

The result is much less readable code where the original intent of the author is harder to recover as opposed to data manipulation in other popular languages (Python, Ruby, etc).

Recommendation: consider using other languages besides bash for heavy data manipulation code.

Absence of non-nullable Types

In some languages there is no straightforward way to specify non-nullable parameters. The programmers are then required to check whether each passed parameter is null. That results more boilerplate code. Let’s look at the following bit of Java code from the popular Apache Flink project:

// flink/flink-java/src/main/java/org/apache/flink/api/java/DataSet.java

protected DataSet(ExecutionEnvironment context, TypeInformation<T> typeInfo) {
    if (context == null) {
        throw new NullPointerException("context is null");
    }
    if (typeInfo == null) {
        throw new NullPointerException("typeInfo is null");
    }

    this.context = context;
    this.type = typeInfo;
}

Asynchronous Computing Model (JavaScript)

In JavaScript for example, progressively more readable code uses:

Again, information loss occurs when programmer’s intention is lost in the code because the code looks like a big struggle against asynchronicity and the language.

Recommendation: prefer async/await over Promises and prefer Promises over callbacks.

Loss of semantic information (JavaScript)

console.log() vs debug('my-module')('my message') in JavaScript. When a programmer chooses to use log() instead of debug(), loss of semantic information occurs. In this case it means more effort in finding the needed information in the output as opposed to simpler turning on and off the relevant debug sections.

Recommendation: use the debug module.

Information Loss at Runtime

Information loss at runtime will manifest as harder debugging.

Empty Catch Clause

This is borderline criminal. Except for very few cases when empty catch clause is really appropriate, by placing empty catch clause in the code, you are setting up a bomb for your colleagues. They will pay with their time, tears and mental health, not to mention they will be hating you. Where is the information loss? At the time the exception is generated, there is useful information about what happened. Empty catch clause loses that information. Result: hard to find exceptions and their causes.

In NGS, there are clear ways to express that you didn’t just forgot to handle the exception (try ... catch(e) { }) but you actually don’t care (or know exactly) what happened:

  • try EXPR without the catch clause at all. If EXPR throws exception, try EXPR evaluates to null, otherwise evaluates to EXPR.
  • EXPR tor DFLT if EXPR throws an exception, evaluates to DFLT, otherwise evaluates to EXPR.

Writing to stdout Instead of stderr

stdout has semantic meaning (result of the computation) and stderr also has semantic meaning (errors description). It will make harder for any wrapper script to deal with a program that outputs errors to stdout or outputs the result to stderr. The semantic information about the text is lost and then needs to be recovered by the caller if the two outputs are mixed.

Wrong exit codes reporting

This one really hinders automation.

if ... then {
    ...
    error("error occurred")
    exit(0) # incorrect error code reported
}

Since it’s easy to forget about exit code, and the common case is that exit() means abnormal termination of the program, in NGS exit() that does not provide an exit code defaults to exit code 1.

Wrong exit codes handling

if [ -e MY_FILE ] ...

This is all over bash scripts… and it’s wrong. Which exit codes [ program/built-in returns? Zero for “yes”, one for “no”, and two for “An error occurred”. Guess what. You can’t handle three distinct cases with two if branches; “An error occurred” is causing the “false” branch of the if to be taken. If you are lucky, you will spot error message on stderr. If you are not lucky, your script will just work incorrectly in some circumstances.

At this point the tradeoff in NGS was made in favor of correctness, not simplicity. if $(test -e MY_FILE) ... in NGS can go three ways: “yes” branch, “no” branch and an exception. After any external process is finished, NGS checks the exit code. For unknown process, non-zero exit code cases an exception. For test and a few others, zero and one are not causing an exception. The exit code checking facility is extensible and one can easily “teach” NGS about new programs.

Broaden your Horizon – Extras

I’ll mention here non-strictly software development related information loss cases.

Untagged Cloud Resources (AWS)

Have you just created an EC2 instance and named it Server or maybe you haven’t tagged it at all? Congratulations, semantic information has just been lost. You colleagues will strugle to understand what is the role of instance.

Recommendation: rigorously tag the resources, have alerts for untagged or improperly tagged resources. In AWS you can also know who created the resource by looking at CloudTrail.

Side note: In Azure, any resource must belong to a “Resource Group” which makes it much easier to track the resources.

GUI

You just performed operation in GUI. The information of what happened was just lost the minute you performed the operation. Good luck reproducing or documenting it.

The plan to combat this in NGS is to have textual representation for each operation that is performed via GUI.

String Concatenation

Every time two strings are concatenated into one, there is some information loss.

Recommendation: instead of parsing unstructured text (result of concatenation) later, consider using structured data format when producing the output. (Example: JSON).


Hope that helps. Have fun!