Seven days for successes, ninety for failures
Check your backend's query logs. Almost every query touches data less than a week old, except the ones from incidents and post-mortems, which look at failures from weeks ago. Two tiers: short for the boring 95 percent, long for errors, slow requests and watched tenants. Leave metrics alone. They are the cheapest way to know something is wrong.
observabilitycost
Longer version: the post this came from.