How to avoid a situation where multiple requests trigger expensive computation simultaneously after cache expiration or absence?