We're currently persisting even empty samples but that "locks" the sample for a collection to something that we know will always produce poor results. We should instead use a similar approach to the one used in the server to decide the sample size based on an acceptable error margin. See https://docs.google.com/document/d/17pEwzppyfEdYDI4GrKv9_Txg9SMxZ2we2h4e8mLVDlA/edit?disco=AAAB6YG2LmE