Between Monday 23rd 15:54 UTC and 15:36 UTC a change was made to make GCP workers poll more quickly for containers and this caused them to pick up more CPU-only containers.

Between 13:48 UTC and 16:18 UTC on September 24th CPU containers were allowed to be scheduled on GCP NVIDIA T4 workers which are the n1 class of GCE VMs.

Both of these changes combined to cause over-scheduling of CPU work onto GCP T4 workers and significantly degrade the performance of the hosted containers (whether CPU-only or T4 GPU-enabled).