-Fix user/project/domain usage buckets being inflated by the number of concurrently running kernels — the fair share aggregator accumulated resource amounts and slice durations separately and multiplied them afterwards, yielding a cross product instead of the sum of per-slice products, so four concurrent GPU kernels reported 4x their real fGPU-seconds. Existing buckets are rebuilt from the unaffected `kernel_usage_records`.
0 commit comments