better errors
This commit is contained in:
parent
78a5df1684
commit
7ab3b7ef36
227
src/_posts/2021-03-20-a-simple-rule-for-better-errors.md
Normal file
227
src/_posts/2021-03-20-a-simple-rule-for-better-errors.md
Normal file
@ -0,0 +1,227 @@
|
||||
---
|
||||
title: >-
|
||||
A Simple Rule for Better Errors
|
||||
description: >-
|
||||
...and some examples of the rule in action.
|
||||
tags: tech
|
||||
---
|
||||
|
||||
This post will describe a simple rule for writing error messages that I've
|
||||
been using for some time and have found to be worthwhile. Using this rule I can
|
||||
be sure that my errors are propagated upwards with everything needed to debug
|
||||
problems, while not containing tons of extraneous or duplicate information.
|
||||
|
||||
This rule is not specific to any particular language, pattern of error
|
||||
propagation (e.g. exceptions, signals, simple strings), or method of embedding
|
||||
information in errors (e.g. key/value pairs, formatted strings).
|
||||
|
||||
I do not claim to have invented this system, I'm just describing it.
|
||||
|
||||
## The Rule
|
||||
|
||||
Without more ado, here's the rule:
|
||||
|
||||
> A function sending back an error should not include information the caller
|
||||
> could already know.
|
||||
|
||||
Pretty simple, really, but the best rules are. Keeping to this rule will result
|
||||
in error messages which, once propagated up to their final destination (usually
|
||||
some kind of logger), will contain only the information relevant to the error
|
||||
itself, with minimal duplication.
|
||||
|
||||
The reason this rule works in tandem with good encapsulation of function
|
||||
behavior. The caller of a function knows only the inputs to the function and, in
|
||||
general terms, what the function is going to do with those inputs. If the
|
||||
returned error only includes information outside of those two things then the
|
||||
caller knows everything it needs to know about the error, and can continue on to
|
||||
propagate that error up the stack (with more information tacked on if necessary)
|
||||
or handle it in some other way.
|
||||
|
||||
## Examples
|
||||
|
||||
(For examples I'll use Go, but as previously mentioned this rule will be useful
|
||||
in any other language as well.)
|
||||
|
||||
Let's go through a few examples, to show the various ways that this rule can
|
||||
manifest in actual code.
|
||||
|
||||
**Example 1: Nothing to add**
|
||||
|
||||
In this example we have a function which merely wraps a call to `io.Copy` for
|
||||
two files:
|
||||
|
||||
```go
|
||||
func copyFile(dst, src *os.File) error {
|
||||
_, err := io.Copy(dst, src)
|
||||
return err
|
||||
}
|
||||
```
|
||||
|
||||
In this example there's no need to modify the error from `io.Copy` before
|
||||
returning it to the caller. What would we even add? The caller already knows
|
||||
which files were involved in the error, and that the error was encountered
|
||||
during some kind of copy operation (since that's what the function says it
|
||||
does), so there's nothing more to say about it.
|
||||
|
||||
**Example 2: Annotating which step an error occurs at**
|
||||
|
||||
In this example we will open a file, read its contents, and return them as a
|
||||
string:
|
||||
|
||||
```go
|
||||
func readFile(path string) (string, error) {
|
||||
f, err := os.Open(path)
|
||||
if err != nil {
|
||||
return "", fmt.Errorf("opening file: %w", err)
|
||||
}
|
||||
defer f.Close()
|
||||
|
||||
contents, err := io.ReadAll(f)
|
||||
if err != nil {
|
||||
return "", fmt.Errorf("reading contents: %w", err)
|
||||
}
|
||||
|
||||
return string(contents), nil
|
||||
}
|
||||
```
|
||||
|
||||
In this example there are two different steps which could result in an error:
|
||||
opening the file and reading its contents. If an error is returned then our
|
||||
imaginary caller doesn't know which step the error occurred at. Using our rule
|
||||
we can infer that it would be good to annotate at _which_ step the error is
|
||||
from, so the caller is able to have a fuller picture of what went wrong.
|
||||
|
||||
Note that each annotation does _not_ include the file path which was passed into
|
||||
the function. The caller already knows this path, so an error being returned
|
||||
back which reiterates the path is unnecessary.
|
||||
|
||||
**Example 3: Annotating which argument was involved**
|
||||
|
||||
In this example we will read two files using our function from example 2, and
|
||||
return the concatenation of their contents as a string.
|
||||
|
||||
```go
|
||||
func concatFiles(pathA, pathB string) (string, error) {
|
||||
contentsA, err := readFile(pathA)
|
||||
if err != nil {
|
||||
return "", fmt.Errorf("reading contents of %q: %w", pathA, err)
|
||||
}
|
||||
|
||||
contentsB, err := readFile(pathB)
|
||||
if err != nil {
|
||||
return "", fmt.Errorf("reading contents of %q: %w", pathB, err)
|
||||
}
|
||||
|
||||
return contentsA + contentsB, nil
|
||||
}
|
||||
```
|
||||
|
||||
Like in example 2 we annotate each error, but instead of annotating the action
|
||||
we annotate which file path was involved in each error. This is because if we
|
||||
simply annotated with the string `reading contents` like before it wouldn't be
|
||||
clear to the caller _which_ file's contents couldn't be read. Therefore we
|
||||
include which path the error is relevant to.
|
||||
|
||||
**Example 4: Layering**
|
||||
|
||||
In this example we will show how using this rule habitually results in easy to
|
||||
read errors which contain all relevant information surrounding the error. Our
|
||||
example reads one file, the "full" file, using our `readFile` function from
|
||||
example 2. It then reads the concatenation of two files, the "split" files,
|
||||
using our `concatFiles` function from example 3. It finally determines if the
|
||||
two strings are equal:
|
||||
|
||||
```go
|
||||
func verifySplits(fullFilePath, splitFilePathA, splitFilePathB string) error {
|
||||
fullContents, err := readFile(fullFilePath)
|
||||
if err != nil {
|
||||
return fmt.Errorf("reading contents of full file: %w", err)
|
||||
}
|
||||
|
||||
splitContents, err := concatFiles(splitFilePathA, splitFilePathB)
|
||||
if err != nil {
|
||||
return fmt.Errorf("reading concatenation of split files: %w", err)
|
||||
}
|
||||
|
||||
if fullContents != splitContents {
|
||||
return errors.New("full file's contents do not match the split files' contents")
|
||||
}
|
||||
|
||||
return nil
|
||||
}
|
||||
```
|
||||
|
||||
As previously, we don't annotate the file paths for the different possible
|
||||
errors, but instead say _which_ files were involved. The caller already knows
|
||||
the paths, there's no need to reiterate them if there's another way of referring
|
||||
to them.
|
||||
|
||||
Let's see what our errors actually look like! We run our new function using the
|
||||
following:
|
||||
|
||||
```go
|
||||
err := verifySplits("full.txt", "splitA.txt", "splitB.txt")
|
||||
fmt.Println(err)
|
||||
```
|
||||
|
||||
Let's say `full.txt` doesn't exist, we'll get the following error:
|
||||
|
||||
```
|
||||
reading contents of full file: opening file: open full.txt: no such file or directory
|
||||
```
|
||||
|
||||
The error is simple, and gives you everything you need to understand what went
|
||||
wrong: while attempting to read the full file, during the opening of that file,
|
||||
our code found that there was no such file. In fact, the error returned by
|
||||
`os.Open` contains the name of the file, which goes against our rule, but it's
|
||||
the standard library so what can ya do?
|
||||
|
||||
Now, let's say that `splitA.txt` doesn't exist, then we'll get this error:
|
||||
|
||||
```
|
||||
reading concatenation of split files: reading contents of "splitA.txt": opening file: open splitA.txt: no such file or directory
|
||||
```
|
||||
|
||||
Now we did include the file path here, and so the standard library's failure to
|
||||
follow our rule is causing us some repitition. But overall, within the parts of
|
||||
the error we have control over, the error is concise and gives you everything
|
||||
you need to know what happened.
|
||||
|
||||
## Exceptions
|
||||
|
||||
As with all rules, there are certainly exceptions. The primary one I've found is
|
||||
that certain helper functions can benefit from bending this rule a bit. For
|
||||
example, if there is a helper function which is called to verify some kind of
|
||||
user input in many places, it can be helpful to include that input value within
|
||||
the error returned from the helper function:
|
||||
|
||||
```go
|
||||
func verifyInput(str string) error {
|
||||
if err := check(str); err != nil {
|
||||
return fmt.Errorf("input %q was bad: %w", str, err)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
```
|
||||
|
||||
`str` is known to the caller so, according to our rule, we don't need to include
|
||||
it in the error. But if you're going to end up wrapping the error returned from
|
||||
`verifyInput` with `str` at every call site anyway it can be convenient to save
|
||||
some energy and break the rule. It's a trade-off, convenience in exchange for
|
||||
consistency.
|
||||
|
||||
Another exception might be made with regards to stack traces.
|
||||
|
||||
In the set of examples given above I tended to annotate each error being
|
||||
returned with a description of where in the function the error was being
|
||||
returned from. If your language automatically includes some kind of stack trace
|
||||
with every error, and if you find that you are generally able to reconcile that
|
||||
stack trace with actual code, then it may be that annotating each error site is
|
||||
unnecessary, except when annotating actual runtime values (e.g. an input
|
||||
string).
|
||||
|
||||
As in all things with programming, there are no hard rules; everything is up to
|
||||
interpretation and the specific use-case being worked on. That said, I hope what
|
||||
I've laid out here will prove generally useful to you, in whatever way you might
|
||||
try to use it.
|
||||
|
Loading…
Reference in New Issue
Block a user