You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
datasets/fake_metric/generate_workloads.py (generated2, #837) filters range windows as ts > END - window AND ts <= END, i.e. (T−d, T]. That is Prometheus 3.0's convention. The bare-metal runs install Prometheus 2.53.2, whose range selectors are closed, [T−d, T]. With 1 s scrapes on whole seconds, a [1m] window reads 60 samples in ClickHouse and 61 in Prometheus, so sum/count/quantile differ and rate extrapolation (first_t, to_start) shifts.
#844's sketch_bench_workloads.py uses the closed window for its classic PromQL and checks it against a Prometheus 2.53 reference.
Fix: change > to >= in select() and extrapolated_rate(). generated2's ts is DateTime (seconds), so the DateTime64 pruning issue #844 works around doesn't apply. Results already collected with generated2 are affected.
Activity