We are using job computes that are spun up as part of a pool. We've also tested with singular jobs/pipelines (very simple table joins that result in dlt tables) and experienced the same issue. So I don't think the issue is high load.