Now, that being said, if you have the ability to set up backups where you can minimize the file system activity, you might be slightly better off doing it, but the ROI is probably fairly low unless it's extremely trivial to set that up for everything that's running.
VSS is a method windows has to allow application to both be informed that a snapshot operation is in progress (and should quiesce their writes) as well as the ability to delay the snapshot operation from occurring (because they are in the middle of a "transaction"). If they delay too much, the snapshot operation will fail and reported back as such to the caller.
edit: also see the former maintainer of VSS at MSFT who responded to me in my other post in this thread
With only a COW filesystem snapshot, the end result will still be correct in case of programs that use transactions (as mentioned, dbs will either have commited all the transactions and the data is saved, or one was in-flight, those will be reverted and you are still in a consistent state). For programs that don’t support transactions both might be faulty, for those that support transactions COW will work correctly and for those that support this API COW will again work and might save a bit fresher version.
Most databases I use keep a write-ahead log, which guards against power failure without requiring an external service. Multiple writes are handled without an issue. If power is lost, the database recovers automatically using whatever is on disk, no matter when it was interrupted.
So a file system snapshot behaves the same and the database can start up from the snapshot even if it was taken at an arbitrary point mid-write.
So now it is quite possible you back up an older WAL coupled with a newer DB file that has partial writes inside. Boom.
This “multi-volume, multi-writer coordination” is the sort of problem VSS solves on windows. I’m not sure there’s anything equivalent on Linux or MacOS that can do multi-volume snapshots consistently.
The term of art is "crash consistent," and any ACID database must preserve all committed state across events such as power loss. Such a database is correctly backed up when copying a simultaneous point in time snapshot across all involved volumes.
Not all databases are truly ACID. Lots of software relies on uncommitted database state. But we're talking about a solved problem here; if you require ACID behavior the means to achieve exactly that are available. Any exception to that statement, including hardware misfeatures or lack of two phase commit across databases, is equivalent to "incorrectly designed."
In a correctly designed system quiescing the database isn't necessary, but might still be used as a precaution or a performance optimization.
unfortuntely most applications aren't like that and assume the filesystem is both atomic and durable
Your figure is too high. Commercial relational databases, many open source databases and widely used file systems are all sufficient. Oracle, SQL Server, DB/2, PostgreSQL and MySql/MariaDB (with InnoDB) are all ACID systems and recover from crash consistent backups fine. Ext4, XFS, NTFS, etc. are all sufficient file systems, unusually in default form. Reliable storage devices are widely available.
Many applications designed around these platforms use COMMIT properly. When they're don't quiescing the database probably won't help with backups anyhow; all you're doing is deferring writes and if the application is not ACID then you can still lose. The database itself, however, will recover and since we're tossing probabilities around, that's 99% of the battle.
meanwhile most of the rest seek randomly in files, buffer and write data without calling ever fsync, or append without a pointer and appropriate write barriers (I'm looking at you journald)
the fact there exist some applications (databases, applications using maildir) that correctly use write barriers doesn't mean all applications are always consistent on disk