Go Error Handling

Errors in Go are values. Treat them as part of the API: choose a strategy per failure mode, propagate with wrapping, inspect with errors.Is/As, and handle each error exactly once.

Core Rules

1. Errors are values, not exceptions. Return them; do not panic across API boundaries.
2. Handle each error exactly once. Either log it or return it — never both.
3. The caller decides what is exceptional. Library code returns; binaries (or top-level handlers) decide whether to log, retry, or exit.
4. Wrap only when you add real context. A wrap that just repeats the function name is noise. Use %w to preserve identity; %v to deliberately hide an unstable type.

Strategy Decision

Pick the simplest strategy that meets the caller's needs:

Read references/strategy-decision.md when the caller's needs are unclear or when migrating between strategies without breaking callers.

Writing Errors

Strings

• Lowercase, no trailing punctuation. Errors are composed: fmt.Errorf("write %s: %w", path, err) reads as one sentence.
• Be specific. "open config: permission denied" beats "failed to open file".
• Do not include the function name. Stack context is added by wrapping at each layer.

Creating

// Opaque — the caller only checks != nil
return errors.New("invalid character in token")

// Sentinel — exported, package-level, named ErrXxx
var ErrNotFound = errors.New("user: not found")

// Typed — when callers need structured fields
type ValidationError struct {
    Field string
    Rule  string
}
func (e *ValidationError) Error() string {
    return fmt.Sprintf("validation: %s violates %s", e.Field, e.Rule)
}

Wrapping and Inspection

Wrap with %w to add context while preserving identity

if err := db.Get(id); err != nil {
    return fmt.Errorf("loading user %d: %w", id, err)
}

Inspect with errors.Is (identity) and errors.As (type)

if errors.Is(err, sql.ErrNoRows) { /* handled */ }

var ve *ValidationError
if errors.As(err, &ve) {
    return reply.BadRequest(ve.Field)
}

Never compare error strings (err.Error() == "...") — strings are not stable API.

Join independent failures

errs := errors.Join(
    validate(name),
    validate(email),
    validate(password),
)
if errs != nil {
    return errs // errors.Is/As walks both branches
}

Read references/wrapping-vs-shadowing.md when deciding between %w (expose) and %v (hide), or when wrapping would leak an implementation detail.

Error Flow

Log or return — not both

// Bad: caller will log it again, producing duplicate lines
if err := svc.Do(ctx); err != nil {
    slog.ErrorContext(ctx, "svc.Do failed", "err", err)
    return err
}

// Good: log only at the boundary that decides the request is done
if err := svc.Do(ctx); err != nil {
    return fmt.Errorf("doing svc work: %w", err)
}

The HTTP handler / job runner / main is the only layer that logs.

Reduce nesting with guard clauses

// Bad
if err == nil {
    if x, ok := f(); ok {
        return x, nil
    }
}
return zero, err

// Good
if err != nil {
    return zero, err
}
x, ok := f()
if !ok {
    return zero, errSomething
}
return x, nil

Panic and Recover

panic is for programmer errors (impossible states) and package initialization. It is never the right way to return a normal failure.

// Acceptable: invariants the type guarantees
func (q *Queue) MustEnqueue(v T) { if err := q.Enqueue(v); err != nil { panic(err) } }

// Acceptable: recover at the goroutine boundary so one bad request cannot kill the server
defer func() {
    if r := recover(); r != nil {
        log.Error("panic recovered", "value", r, "stack", debug.Stack())
        http.Error(w, "internal error", 500)
    }
}()

Do not use recover to convert panics into errors as normal control flow.

For custom error types, implement Unwrap() error (or Unwrap() []error in Go 1.20+) so errors.Is/As can reach the cause. See references/wrapping-vs-shadowing.md.

Anti-Patterns

Verification Checklist

Before finishing an error-handling change:

• [ ] No err.Error() string comparisons
• [ ] All wrapping uses %w (or %v is intentional and commented)
• [ ] Functions return the error interface, not concrete types
• [ ] Each error is logged at most once (at the request/job boundary)
• [ ] Sentinels are package-level var ErrXxx = errors.New(...)
• [ ] Typed errors expose only the fields callers actually need

References

• references/strategy-decision.md — picking opaque vs sentinel vs typed
• references/wrapping-vs-shadowing.md — %w vs %v decisions
• references/typed-nil-trap.md — why returning *MyErr breaks == nil