How to save hardware for production machine learning

There are dozens of services with ML models in production at Avito, and Oleg's team has long experience of saving hardware for them. Oleg has some insights about how to save up to 200% servers, and he is ready to share them (insights, not servers).