Change checkpoint_data and restore_connection_on_error to subtransaction checkpoint_data and restore_connection_on_error used to commit() and rollback() the db connection. This causes additional I/O and load. Changed them to use 'SAVEPOINT name' and 'ROLLBACK TO name' to get a faster method for handling errors within a tranaction. One thing to note is that postgresql (unlike SQL std) doesn't overwrite an older savepoint with he same name. It keeps all savepoints but only rolls back to the newest one with a given name. This could be a resource issue. I left a commented out release statement in case somebody runs into an issue due to too many savepoints. I expect it to slow down the import but....

[Unit]
Description=Roundup Server - roundup

[Service]
User=roundup
WorkingDirectory=/path/to/roundup/trackers/roundup
SyslogIdentifier=roundup-server-roundup
ExecStart=/path/to/gunicorn -b unix:///var/run/roundup/roundup.sock wsgi -w 4 --preload --max-requests 128 --max-requests-jitter 32 --timeout 60 --graceful-timeout 60 --access-logfile - --error-logfile
ExecReload=/bin/kill -HUP $MAINPID
ExecStop = /bin/kill -s TERM $MAINPID
KillMode=mixed
Restart=on-failure
RestartSec=15s

[Install]
WantedBy=multi-user.target