Is vast.ai reliable for production inference where downtime is important?