note 1 min

A small pattern for wrapping errors that survives errors.Is

Sentinel errors and context are not in tension, but you have to wrap in the right direction to get both.

Two things I want from an error: enough context to know where it came from, and enough structure that a caller can branch on it. It's easy to get one and lose the other.

var ErrNoBackend = errors.New("no healthy backend")

func (p *Pool) Pick(key uint64) (*Backend, error) {
	b := p.table.Lookup(key)
	if b == nil {
		return nil, fmt.Errorf("pool %q: %w", p.name, ErrNoBackend)
	}
	return b, nil
}

The %w is the whole trick. errors.Is(err, ErrNoBackend) still matches at any depth, and the message still tells you which pool.

Where I see this go wrong is %v instead of %w — it reads identically in logs and silently breaks every caller doing errors.Is. That's a one-character bug with no symptom until someone's retry logic quietly stops retrying.

If you're joining multiple failures, errors.Join preserves matching across all of them, which is what you want for something like health checks where several backends failed for different reasons.

#go #errors