Go error handling

Go Error Handling

In Go, errors are values — they are created by code and consumed by code.

Choosing an Error Strategy

1. System boundary (RPC, IPC, storage)? → Wrap with %v to avoid leaking internals
2. Caller needs to match specific conditions? → Sentinel or typed error, wrap with %w
3. Caller just needs debugging context? → fmt.Errorf("...: %w", err)
4. Leaf function, no wrapping needed? → Return the error directly

Default: wrap with %w and place it at the end of the format string.

---

Available Scripts

• scripts/check-errors.sh — Detects error handling anti-patterns: string comparison on err.Error(), bare return err without context, and log-and-return violations. Run bash scripts/check-errors.sh --help for options.

---

Core Rules

Never Return Concrete Error Types

Never return concrete error types from exported functions — a concrete nil pointer can become a non-nil interface:

// Bad: Concrete type can cause subtle bugs
func Bad() *os.PathError { /*...*/ }

// Good: Always return the error interface
func Good() error { /*...*/ }

Error Strings

Error strings should not be capitalized and should not end with punctuation. Exception: exported names, proper nouns, or acronyms.

// Bad
err := fmt.Errorf("Something bad happened.")

// Good
err := fmt.Errorf("something bad happened")

Return Values on Error

When a function returns an error, callers must treat all non-error return values as unspecified unless explicitly documented.

Tip: Functions taking a context.Context should usually return an error so callers can determine if the context was cancelled.

---

Handling Errors

When encountering an error, make a deliberate choice — do not discard with _:

1. Handle immediately — address the error and continue
2. Return to caller — optionally wrapped with context
3. In exceptional cases — log.Fatal or panic

To intentionally ignore: add a comment explaining why.

n, _ := b.Write(p) // never returns a non-nil error

For related concurrent operations, use errgroup:

g, ctx := errgroup.WithContext(ctx)
g.Go(func() error { return task1(ctx) })
g.Go(func() error { return task2(ctx) })
if err := g.Wait(); err != nil { return err }

Avoid In-Band Errors

Don't return -1, nil, or empty string to signal errors. Use multiple returns:

// Bad: In-band error value
func Lookup(key string) int  // returns -1 for missing

// Good: Explicit error or ok value
func Lookup(key string) (string, bool)

---

Error Flow

Handle errors before normal code. Early returns keep the happy path unindented:

// Good: Error first, normal code unindented
if err != nil {
    return err
}
// normal code

Handle errors once — either log or return, never both:

Error encountered?
├─ Caller can act on it? → Return (with context via %w)
├─ Top of call chain? → Log and handle
└─ Neither? → Log at appropriate level, continue

Read references/ERROR-FLOW.md when structuring complex error flows, deciding between logging vs returning, implementing the handle-once pattern, or choosing structured logging levels.

---

Error Types

Advisory: Recommended best practice.

Caller needs to match?	Message type	Use
No	static	errors.New("message")
No	dynamic	fmt.Errorf("msg: %v", val)
Yes	static	var ErrFoo = errors.New("...")
Yes	dynamic	custom error type

Default: Wrap with fmt.Errorf("...: %w", err). Escalate to sentinels for errors.Is(), to custom types for errors.As().

Read references/ERROR-TYPES.md when defining sentinel errors, creating custom error types, or choosing error strategies for a package API.

---

Error Wrapping

Advisory: Recommended best practice.

• Use %v: At system boundaries, for logging, to hide internal details
• Use %w: To preserve error chain for errors.Is/errors.As

Key rules: Place %w at the end. Add context callers don't have. If annotation adds nothing, return err directly.

Read references/WRAPPING.md when deciding between %v and %w, wrapping errors across package boundaries, or adding contextual information.

Validation: After implementing error handling, run bash scripts/check-errors.sh to detect common anti-patterns. Then run go vet ./... to catch additional issues.