See the linked BFs for more details.

Context: in PERF-3361 we found that much of the noise was due to noisy bucketing patterns in the underlying data. A recommendation here is to look into whether we can have the data population phase of these workloads restore a pre-bucketed archived dataset for maximum repeatability.

I think this will involve some improvements to the genny tooling to be able to import a dataset - but it's possible we can just use what DSI has and do this setup code before the workload starts. Leaning on DSI to do the setup will just make it harder for a local developer to reproduce, since it's not as self-contained. But we could combat this with documentation.