In general, all data of live clusters will also be stored locally at a PageServer.
Only in recovery scenarios will a PageServer not hold the data that is needed to serve the requests of a compute node - but that would recover quickly as the local cache of the PageServer is repopulated with data from S3.
Phrased another way, would a query that needs to access a relatively large amount of data (10-100 GB) ever need to read from s3, incurring extra latency?