We may have found the culprit. Linear models have a different way of controlling parallelism which we did not account for. We're going to disable linear models until we patch the backend since they affect everything else that runs on the same server. I'll post something specific shortly.