What did you do?

Run Prometheus with out-of-order support enabled

What did you expect to see?

prometheus_tsdb_sample_ooo_delta to have the right buckets filled.

What did you see instead? Under which circumstances?

prometheus_tsdb_sample_ooo_delta is initialized with buckets as seconds

But Prometheus uses milliseconds as its timestamp. So observing the ooo_delta with milliseconds will lead to wrong buckets being filled. See:

Until Prometheus starts using native histograms, TSDB can compromise from its timestamp unit agnostic design for this metric and do float64(delta)/1000 instead.

This issue is broken down from #11329