A strong answer. “There is no free answer here, and I would say that first — every option discards correct data or changes a published number, and the job is to choose deliberately rather than to have the framework choose for me.

Concretely there are three. Drop them beyond a watermark, which keeps the numbers stable and means we are knowingly slightly wrong, and the size of the error is measurable if I count what I dropped. Hold the window open longer, which costs state proportional to the lateness allowance for every window, not just this one — a four-hour allowance on a one-minute window is a very different system. Or accept restatement: publish, then correct, which is honest and requires that every consumer downstream can tolerate a number changing, which most dashboards and no commission calculation can.

What I would actually do depends on the consumer. For operational alerting, drop and count. For finance, restate — they already understand restatement and would rather be correct late than stable and wrong. For a general dashboard, a bounded allowance plus a visible ‘as of’ timestamp.

And I would instrument the lateness distribution rather than guessing, because the allowance should come from data. Usually the tail is much longer than anyone assumes and the ninety-fifth percentile is a far better basis than a round number someone picked.”