The checkpoint file contains a bitmap of all row groups at that time. A single full load is used during loading, and cache eviction is implemented using a lease mechanism. The cache is handled locally by trino worker nodes to avoid load bottlenecks on the retina nodes.