If we don't want to use any of the preceding options that were outlined, our last resort is to copy the raw files using cp/rsync, or something equivalent. This is generally not recommended, for the following reasons:
- We need to stop all of the writes before copying files
- The backup size will be larger, since we need to copy indexes and any underlying padding and fragmentation storage overhead
- We cannot get point-in-time recovery by using this method for replica sets, and copying data from sharded clusters in a consistent and predictable manner is extremely difficult
Making a backup by copying raw files should be avoided, unless no other option really exists.