How is Zingg implemented at scale for C360 customer matching and golden record resolution in production?

Hello, I am just investigating Zingg for an existing customer. Do we have any existing patterns that you can share regarding how Zingg is actually implemented in production? The customer would like to use something like Zingg for their C360 initiative and they are unsure of how the model is deployed... how it matches and resolved the "golden record" at inference time. Does it need to cached the entire master (previously loaded and cleansed) record set each time a new record is presented to the model at inference time..in order to make a match? Does it need to scan potentially 500,000 customer records each time it trys to resolve a match when a new record is presented to the model? I'm just trying to understand how this is operationalized at scale for incremental feeds of new customer data. BTW: This is a databricks deployment where the sources could be amongst 10 different source systems presenting this customer data. Any insight you could provide would be greatly appreciated.