At $work we properly rinse autoscale shared project runners. (we make lots of CUDA lumps so we need GPUs to test, and they are uber expensive.)

I wasn't aware of multi-pipeline, so if I bump into that problem again, I'll give it a go