I've released a Go linter to check for the rule described in this post. You can find it here: wrapcheck

This is a small post covering error wrapping in Go that will build on the work from Dave Cheney.

So recently, after spending more time than I should hunting down the source of an error, I sparked the discussion within my team at work about how we can cut down on the time to debug, specifically, by making sure that errors aren't handed back to the caller without forgetting to attach extra information.

We already have a pretty good practice in place to wrap errors with additional info before we hand them up the stack, however this was missing in this particular case.

Essentially, we were seeing something akin to:

time="2020-08-04T11:36:27+02:00" level=error error="sql: no rows in result set"

...not particularly useful.

To help with debugging here, it's important that additional information is attached to the error to give the developer more context.

Extending errors

By now you've probably already heard the popular Go proverb of:

Don’t just check errors, handle them gracefully

Handling an error in this case would be to add this additional information and pass it back up the stack. However, we're not just limited to that. If we take a closer look at what an error is in Go, we see the following interface.

type Error interface {
  Error() string
}

Very simple, yet very powerful. However this power comes with many possibilities, and unfortunately many possibilities comes with no standard way of handling things. Ultimately, it is up to you to define what form of extension is most suitable for your application.

An example is in Rob Pike's Upspin project. They use a custom error struct which looks like:

type Error struct {
  Path upspin.PathName
  User upspin.UserName
  Op   Op
  Kind Kind
  Err  error
}

You can read more about how exactly they use this error definition in a blog post they published.

This is a bit intensive for many programs however, so we'll take a look at a less intense, more widely used method for adding information to errors, called wrapping.

Wrapping errors prior to Go 1.13

Go has had error wrapping since the introduction of the errors package was introduced in early 2016.

This shortly followed with the errors.[New|Errorf|Wrap|Wrapf] methods implementing the interface:

type Stack interface {
  Stack() []uintptr
}

The functionality for adding stack traces to existing errors came with the introduction of the errors.WithStack() method on the same package.

This returns an error with a format method which will print the stack trace when used with the %+v verb.

import "github.com/pkg/errors"

...

func (db *DB) getTansactionByID(tranID string) (Transaction, error) {
  sql := `SELECT * FROM transaction WHERE id = $1;`

  var t Transaction
  if err := db.conn.Get(&t, sql, tranID); err != nil {
    return Transaction{}, errors.Wrap(err, "failed to get transaction")
  }

  return t, nil
}

...

Which, when logging the error with the %+v verb, produces something along the lines of:

sql: no rows in result set
main.main
  /Users/tom/Documents/tmp/main.go:10
runtime.main
  /usr/local/Cellar/go/1.14.6/libexec/src/runtime/proc.go:203
runtime.goexit
  /usr/local/Cellar/go/1.14.6/libexec/src/runtime/asm_amd64.s:1373

We can use this to our advantage when creating errors to include a stack trace in them. This can then be logged along with the error in order to aid locating the source of the error.

One downside of this approach is the performance penalty. The necessary call into runtime.Stack(buf []byte, all bool) int incurs a non-negligible cost, and should be avoided in hot paths.

Stack traces are also not particularly human friendly. They will only identify where the error occurred, not why. Therefore, it's still good to add extra context to your errors.

Another minor nitpick is that when wrapping errors multiple times with github.com/pkg/errors, it will leave you with multiple stack traces, one for each time you wrap the error. This creates a rather large set of logs when you finally print the error if your stack is deep 🃏.

Wrapping errors in Go 1.13

In version 1.13, Go introduced the %w verb, as well as a few methods on the errors package to support error identification and incorporate wrapping into the standard library.

By using fmt.Errorf() with the %w verb, we're able to enhance errors with additional information using only the standard library, while allowing them to remain inspectable should we need to identify the underlying cause of an error.

It's important to keep in mind however, when wrapping with %w, you are implicitly exposing the error to consumers of your package. Only use this when you plan to support the error type and avoid exposing implementation details. You can use the %v verb instead for non-inspectable errors (no errors.Is(), errors.As()).

Now having fmt.Errorf() available is good, however if you never use it in your programs you're still going to end up with logs which are hard to decipher. Now you could most certainly wrap every error returned at every point of your program. However, you will end up with a lot of redundant wraps which don't add a whole lot of value and reduce the signal-to-noise ratio in your logs.

Hence, it would be helpful to have a simple set of rules to follow for when errors should be wrapped. This is by no means an exhaustive set, however should cover the majority of common cases within most programs.

  1. When you have additional context, which you think would be useful to give to the developer reading the log. This is important, as rules are not infallible.
  1. Errors returned from another package. Wrapping these helps identify the entry point of the error into your program. e.g.
...

func (db *DB) getTansactionByID(tranID string) (Transaction, error) {
  sql := `SELECT * FROM transaction WHERE id = $1;`

  var t Transaction
  if err := db.conn.Get(&t, sql, tranID); err != nil {
    return Transaction{}, fmt.Errorf("failed to get transaction with ID %s: %v", tranID, err)
  }

  return t, nil
}

...
  1. Errors returned by interface methods. Interfaces may be implemented by a separate package, therefore wrapping the errors here with context may be valuable.

This set of rules gives a good starting point for helping you identify where in your program you should wrap errors. In order to help out with this, I've written a linter to identify areas where your code is not wrapping errors and it possibly should be.

Wrapcheck

Wrapcheck is a new linter for Go which I built in order to maintain consistency across a codebase with regards to point (1) and (2).

For example, the following code which calls into the sql package will be reported as an error as follows.

func (db *DB) getUserByID(userID string) (User, error) {
  sql := `SELECT * FROM user WHERE id = $1;`

  var u User
  if err := db.conn.Get(&u, sql, userID); err != nil {
    return User{}, err // wrapcheck error: error returned from external package is unwrapped
  }

  return u, nil
}

func (db *DB) getItemByID(itemID string) (Item, error) {
  sql := `SELECT * FROM item WHERE id = $1;`

  var i Item
  if err := db.conn.Get(&i, sql, itemID); err != nil {
    return Item{}, err // wrapcheck error: error returned from external package is unwrapped
  }

  return i, nil
}

The fix is to add wrapping to the errors before returning them. The linter however is a bit more flexible, and doesn't mind if you still want to use pkg/errors. It will be satisfied as long as the error is not returned bare.

The linter was written with the new(ish) go/analysis tooling, which greatly simplifies AST traversal as well as setup and testing.

The linter has a fair number of tests. You can have a look at all the specified cases in the testdata directory.

Please feel free to use it in your own projects and report any issues you may come across. As mentioned, the linter is meant to handle gracefully the standard case, there are most certainly cases which slip past it. But feel free to file an issue for these as well.