[dom0]

Yes, that's right, concurrent addition of the same chunks would generally mean that some work is wasted; so concurrent long running jobs would not synchronize well in this model, and lock-free performs clearly better there.

The only operation which inherently has to be guarded by a lock in Borg is inserting the archive pointer into the manifest (root object, see https://borgbackup.readthedocs.io/en/latest/internals/data-s...). I suppose it would be possible to work around that without locking or to use the usual hacks around EC, put/get/check/get/check?put/get/check?put etc. until it's "probably there".

Deleting / pruning archives would still require a full lock due to the same conceptual issues that your two-phase GC avoids. The same goes for "check".