james mckay dot net
because there are few things that are less logical than business logic

Posts tagged: exceptions

To throw or not to throw? Non-exceptional error conditions

I’ve said before that the general advice that “you should only throw exceptions in response to exceptional conditions” is bad advice, because it is ambiguous and unclear as to what exactly constitutes an exceptional condition. But nevertheless, there is a valid point behind it.

The point is that exceptions are (a) slow, and (b) painful to step through in the debugger.

This means that if there are error conditions that crop up frequently, you will want to handle them without an exception being thrown wherever possible.

The classic example here is, of course, attempting to parse an integer. This is why .NET gives you the Int32.TryParse method. If it is provided with an invalid input string, it returns false instead of throwing an exception.

Now there are two important points to note about Int32.TryParse.

First: it returns false only in response to a single, specific error condition: namely, bad user input. Any other failure mode should still throw an exception. This is the case with the overload of Int32.TryParse that also takes some localisation and formatting parameters, for example.

Second: its name tells us that it returns an error code in response to this specific condition. If it didn’t have that Try... prefix, it would, as Yoda said, be a case of “Do or do not. There is no try.” This means that the meaning of exceptions — that the method can not do what its name says that it does — is preserved.

Third: it is provided alongside an equivalent method, Int32.Parse, that throws an exception as normal. This gives consumers of your API a choice about which method to use.

Fourth, and most importantly, it is not implemented by throwing and re-catching exceptions. If you write a method like this, you are doing it wrong:

public bool TryParse(string input, out int output)
    try {
        output = Int32.Parse(input);
        return true;
    catch {
        return false;

Rather, you should do it the other way round:

public int Parse(string input)
    if (input == null) {
        throw new ArgumentNullException("input");
    int result;
    if (Int32.TryParse(input, out result)) {
        return result;
    else {
        throw new FormatException("The parameter 'input' was not in the correct format.");

Why should this be? Remember that exceptions are expensive both in terms of execution time and stepping through them in the debugger. If you are implementing Try... methods by wrapping other methods in a try ... catch block, you will be losing the benefits of having a Try... method in the first place.

Recommendation: Use this pattern whenever you are taking user input that may or may not be valid. But do not use this pattern for error conditions that are not a direct result of invalid user input.

To throw or not to throw? Doing something that has already been done

The general rule for throwing exceptions (or, in Go, returning an error code) is that your function could not do what its name says that it does. This is pretty clear-cut most of the time, but there are corner cases where it’s ambiguous as to whether your function was actually successful or not.

Attempting to do something that has already been done.

Is it an error condition when your method is trying to do something that has already been done? You could argue it both ways. On the one hand, your method hasn’t been able to do what it was asked to, so it could be regarded as an error. On the other hand, the outcome of your method is what you wanted anyway, so in that sense it isn’t.

Returning success in such a condition is called idempotency. It is convenient, since it allows you to just call the method without including a superfluous and unnecessary check, and personally, it’s the option that I would prefer.

Unfortunately, there doesn’t seem to be a clear consensus on this one.

The task of creating a directory (together with any missing parents) in your filesystem is a case in point. By default, Python’s os.makedirs function throws an OSError exception, though Python 3.2 added an extra argument, exist_ok, to allow you to specify idempotent behaviour.

On the other hand, in .NET, the System.IO.Directory.CreateDirectory method takes the idempotent approach. If the directory already exists, it does not throw an exception, but merely returns a DirectoryInfo instance for the existing directory.

What was done previously was not exactly what you wanted.

Unfortunately, it isn’t always that clear cut, and things can get a bit messy. What happens, for example, if you have asked your method to create a directory with some security rules that don’t match what’s already there? Both Python and .NET (in later versions of the .NET Framework, though not in .NET Core) give you this option.

In .NET, the method simply returns without doing anything. Python 3.4.1 or later does the same when the exist_ok argument is set to True. However, this was not always the case. Python 3.2 through 3.4.0 threw an exception in this case. In theory, this should have been the correct thing to do, but unfortunately it introduced a race condition and it was not possible to fix it safely, so this was removed.

Recommendation: Always check the documentation for any methods that you call, because there is not much in the way of consistency here. But if your own function is seeking to do something that has already been done, return success by default.

Positive, negative, or error?

In my recent blog post about error handling in Go, I wrote about why I thought that they made the wrong decision by not implementing exception handling. One of the posts I linked to was this article by a Go developer called Dave Cheney.

In another article by the same author, he says that he will prove that Go’s error handling is superior. However, the example he gives is based on flawed logic. Nevertheless, it’s instructive to look at it, because there are some important things that we can learn from it.

His argument is based around this snippet of code:

package main
import "fmt"
// Positive returns true if the number is positive, false if it is negative.
func Positive(n int) bool {
    return n > -1
func Check(n int) {
    if Positive(n) {
        fmt.Println(n, "is positive")
    } else {
        fmt.Println(n, "is negative")
func main() {

which, as he points out, gives this answer:

1 is positive
0 is positive
-1 is negative

which is wrong, because 0 is not positive. However, as he says, 0 is not negative either. To fix this, he proposes returning an error condition as well:

// Positive returns true if the number is positive, false if it is negative.
// The second return value indicates if the result is valid, which in the case
// of n == 0, is not valid.
func Positive(n int) (bool, bool) {
    if n == 0 {
        return false, false
    return n > -1, true
func Check(n int) {
    pos, ok := Positive(n)
    if !ok {
        fmt.Println(n, "is neither")
    if pos {
        fmt.Println(n, "is positive")
    } else {
        fmt.Println(n, "is negative")

Here’s the flaw.

What a function does should be defined by its name and any relevant established conventions.

The flaw in Dave’s logic is that he is trying to use the Positive() function for purposes for which it should not be intended. There is nothing in the function’s name that tells you that it will determine whether or not a number is negative. It only tells you that it will determine whether or not it is positive.

You can see this more clearly if we change the requirements a bit. What happens if he was asked to produce a program that, rather than telling whether a number is positive or negative, would tell whether it was prime or Fibonacci? The series of prime numbers goes 2, 3, 5, 7, 11, 13, 17, 19 … whereas the series of Fibonacci numbers goes 1, 1, 2, 3, 5, 8, 13 and so on. But should we have functions IsFibonacci() and IsPrime() that throw errors for 4, 6, 9, 10, 12, 14, 15, 16, 18, 20? Of course not!

What he needs to do instead is declare a second function, Negative():

func Positive(n int) (bool) {
    return n > 0
func Negative(n int) (bool) {
    return n < 0
func Check(n int) {
    pos := Positive(n)
    neg := Negative(n)
    if pos {
        fmt.Println(n, "is positive")
    } else if neg {
        fmt.Println(n, "is negative")
    } else {
        fmt.Println(n, "is neither")

Neither a function called Positive() nor one called Negative() should have any preconditions whatsoever. A number — any number, zero included — is either positive or it isn’t. Zero is not positive, so you would return false. If we were dealing with floating point numbers, NaN (not a number) would not be positive, so you would return false. Strictly speaking, in dynamically typed languages, null, "Hello world", an HTTP client class, or an aardvark, are not positive either, so you would return false.

Remember that, whether you are using exceptions or error codes, an error indicates that your function could not do what its name says that it does. Making zero — or anything else — an error condition violates this rule, is outside of the scope implied by the function’s name and any well known conventions I can think of, and as such, it is counterintuitive, confusing, and wrong.

Error handling is the one thing that puts me off learning Go

The Go-pher

It seems that all the cool kids are learning Go these days. It’s certainly appealing to be able to get the performance of C without all the headaches, and to be able to package your program up into a single, tight binary without masses and masses of bloated dependencies. Furthermore, since it’s the language of choice for a lot of important software, such as Kubernetes and Terraform, I’m probably going to have to get my head round it one way or another sooner or later.

But there’s one thing about Go that I really, really do not like one little bit: its approach to error handling. Rather than having exceptions, it reports errors as return codes.

Go’s justification for not having exceptions is as follows:

We believe that coupling exceptions to a control structure, as in the try-catch-finally idiom, results in convoluted code. It also tends to encourage programmers to label too many ordinary errors, such as failing to open a file, as exceptional.

Go takes a different approach. For plain error handling, Go’s multi-value returns make it easy to report an error without overloading the return value. A canonical error type, coupled with Go’s other features, makes error handling pleasant but quite different from that in other languages.

This is an unjustified assertion. They give no explanation as to what is supposed to be convoluted about exception handling code, nor do they give any examples. Consequently, I have no idea what on earth they are talking about, and neither do you. Sure, I’ve seen code that gets exception handling wrong, but that was more due to the code itself being bad rather than any problem with the concept of exceptions itself. I’ve also worked with codebases that get it right, and all I can say is that exceptions done right are much easier to work with and reason about than error codes.

Wrong reasons for objecting to exceptions

There are various reasons why people don’t like exceptions. Some people react against them because they were popularised by Java, and a whole lot of other Bad Things were popularised by Java as well. And yes, maybe Java did make some mistakes by implementing checked exceptions, but please, don’t throw out the baby with the bathwater.

Others complain about exceptions because people get them wrong, doing stupid things like this:


Please, people. The correct response to misuse is not disuse, but proper use. People will do stupid things with any programming language construct. It doesn’t mean that the constructs themselves are bad.

Others consider not having to write extra error handling code as laziness. But so what? Work is not about favouring busyness over laziness; work is about delivering value to your customers. If you can deliver the same value in half the time with half the bugs and half as many lines of code, you’re not being lazy; you’re being efficient.

Others complain about exceptions crashing their Python or C# code with an ugly looking stack trace. But this is easy to fix, simply by implementing a global exception handler at the top of your code, sending the stack trace to a logging service such as ElasticSearch, and just showing an appropriate message. In any case, which is worse — a stack trace, or a program that silently corrupts your data?

Others complain that they mean that you don’t know which functions might throw errors and which might not. But the only safe assumption that you can make is that any line of code might throw an error — and often for reasons that you are not expecting and did not anticipate.

What are you supposed to do with errors anyway?

The most important thing you need to realise is that both exceptions and error codes should have a very specific meaning — namely, that your method was unable to do what its specification says that it does. It could have failed for any number of reasons: bad user input, missing dependencies, external services having gone offline, timeouts, foreign key violations, null references, division by zero, out of memory, stack overflow, array bounds errors, or even literal bugs. But the important information that they convey is that you asked some other code to do X, and for whatever reason, it did not do X.

It is completely unhelpful to try to categorise them as “ordinary errors” or “exceptional errors.” This distinction is so vague and ambiguous as to be effectively meaningless, and in any case will depend more on context and your own use cases than on any intrinsic properties of the exceptions themselves. The main distinction that you need to make with errors is between those that you have anticipated and can meaningfully correct, and everything else.

For errors that you are able to handle, the correct action will usually be specified in your user stories, and as such, they will need to be handled on a case by case basis. But for errors that you have not yet anticipated, 95% of the time the correct action will be to assume that your own code is also unable to do what it is supposed to, stop what it is doing, and report a failure to the caller.

It is almost never appropriate for your code to carry on regardless after an error. If it does so, it will be running under assumptions that are incorrect. At best, it will result in further errors. At worst, it will silently corrupt your data.

Exceptions are about convention over configuration and safe defaults

Convention over configuration is a principle of language and framework design that says that if one specific course of action predominates, it should be made an implicit convention, and extra code should only be needed to override it.

With error codes, the default behaviour is to do precisely what you are not supposed to do — carry on regardless. Consequently, every single function call needs to be followed by a test for the return value. And the code to do so will be mindlessly, frustratingly repetitive:

    if err := datastore.Get(c, key, record); err != nil {
        return &appError{err, "Record not found", 404}
    if err := viewTemplate.Execute(w, record); err != nil {
        return &appError{err, "Can't display record", 500}

But what this is doing is exactly what exceptions do anyway! The whole point of exceptions is to take this repetitive boilerplate code and make it implicit. Other mechanisms, such as try/catch/finally blocks, exist to provide clear and specific ways to override this convention. The result is code that is clearer and easier to understand, with a significantly improved signal-to-noise ratio.

Yet the language designers of Go consider this boilerplate a virtue!

In Go, error handling is important. The language’s design and conventions encourage you to explicitly check for errors where they occur (as distinct from the convention in other languages of throwing exceptions and sometimes catching them). In some cases this makes Go code verbose, but fortunately there are some techniques you can use to minimize repetitive error handling.

Whatever happened to DRY? Whatever happened to convention over configuration?

Cleaning up

When your exception handling code demands more complex scenarios than just propagating the error up the call stack, 90% of the time it will simply be to clean up: close file handles, release locks, roll back transactions, and then propagate the error condition up the call stack. Furthermore, the cleanup code will usually be common to a whole block of other method calls, any of which could raise an error. This Python code is an example:

f = open('data.json')
    transaction = repository.start_transaction()
    data = json.load(f)
    repository.update(data, transaction)

I shall leave it as an exercise for the reader to translate this block of code into Go. Now Go gives you the defer instruction to allow you to queue up functions to run at the end of your method, but if you have to run separate code for success (transaction.commit()) and failure (transaction.rollback()), as is the case here, your code will be significantly more complex. Additionally, many exception-based languages give you syntactic sugar for these most common exception handling cases — in particular, using in C#, or with in Python.

Stick to the conventions of your language

Of course, if you’re using Go, handling error codes is what you have to do. The Go designers decided to do without exceptions, and to introduce them at this stage would just cause confusion. You would end up with some functions returning error codes while others throw exceptions under exactly the same failure conditions.

Now Go does have a panic/recover construct that is similar to exceptions in some respects. But it is rarely — and inconsistently — used. We are told that it is supposed to be reserved for “truly exceptional conditions.” But what, exactly, makes one condition “truly exceptional” and another not? Why, for example, are array bounds errors exceptional, but Println errors, bad format strings, and broken connections are not? There is neither rhyme nor reason to the distinction.

The Go community seems to be avoiding problems with error handling for now, but that is mainly because Go programmers tend to be experienced, high-end developers who are used to the discipline of meticulously writing all the extra error-handling code. But with the rise in Go’s popularity, sooner or later it is going to experience an eternal September, with newbies piling in and forgetting the all-important error-handling boilerplate code left, right and centre, and when that happens, they will discover that return codes instead of exceptions are no panacea.