Filestore is broken on object pushes
This is probably a bug with transaction handling rather than pushes specifically.
Steps to reproduce:
1) Start a cluster (our testing was with 4 nodes)
2) Write some objects (ie, with rados bench 60 write)
3) Adjust replication level from 2 to 3.
4) Notice that scrubs turn up mis-matches and that new replicas don't have the objects they should.
We don't know exactly when this was introduced, but presumably it was on the lfn merge or the transaction error-handling rollback code.