Several years ago, IRSA embarked on the task of regularizing all our image and spectral metadata by converting them to the CAOM data model. Our goals are to reduce the number of custom tools needed to support data ingestion and user interfaces, standardize our internal and external interfaces on VO services and protocols, and enable ObsCore/ObsTAP data access. We will describe the progress we've made, the hurdles we're still struggling with, and lessons learned about how to approach changing the data model layer underneath a petabyte-scale (and growing) data archive spanning dozens of diverse datasets.