Cost control

- [Instructor] While running Machine Learning…with Google Cloud Platform provides…unlimited scalability and performance, it comes at a cost.…GCP resources are built based on usage…for both training and predictions.…Enthusiastic engineers might be running…a number of model trials,…but they can quickly run up the bill.…So, how best to optimize cost?…Choose a type of machine or TPU based on the work involved.…

Real-time predictions might need faster response times,…and hence better resources.…Online predictions, where the prediction function…is called for every input record,…is much for expensive than doing batch predictions,…where a list of import records are provided.…Plan for resource usage, and monitor them appropriately.…Every model in GCP provides performance metrics,…so monitor them periodically,…and carefully optimize as far as costs.…

When trials and model building need to be done,…try to do them on local resources,…like your laptop or just a compute engine.…Expand the size of training and test datasets…only after confident results are obtained…