How it works
File transfer and backups
How SFTP logins are jailed so no account sees another, why FTPS ships switched off, what the backups take, and the one thing nothing re-checks after install.
File transfer is served by the host's own OpenSSH. This project wrote no daemon for it.
Each login is jailed, and the jail is real
A login is chrooted into a root-owned jail with the account's real home bind-mounted inside it. This has two consequences:
- Nothing about the home's ownership had to change to make the jail work, so the account still owns its own files exactly as it did.
- No other account is visible from inside. The customer is not asked to stay in the jail's directory. It is the only directory there is.
The login carries the account's own uid, so an uploaded file belongs to the account exactly as one
it created itself. It gets internal-sftp and no shell. This is tested by a real session being
refused a command, not by checking that a directive is in a file.
Port forwarding is off. A credential that can forward ports would be a tunnel into the server's private network.
Resetting or deleting a login ends its open sessions
A login's password and its public keys are checked when a connection is made, and never again for a session the server has already admitted. So changing a credential does not cut off someone who is already connected with the old one. When you reset a login's password, the panel ends every session that login has open. When you delete a login, it ends them before removing the login. An upload that is in flight is cut at whatever byte it had reached, and the partial file stays. Only that login's sessions end; the account's other logins are not touched. If the server cannot confirm the sessions ended, a delete is refused and can be tried again, and a reset still returns the new password but records the failure in the audit log as SFTP sessions ended. That entry tells you the old sessions may still be open.
Public keys live outside the jail
The same block says where an SFTP login's public keys are read from:
/etc/ssh/maran-sftp-keys/<login>, one root-owned file per login. A key a customer puts in their
own home is never read, because a customer must not be able to write the file that decides who may log in.
The server agent writes that file only from a checked list. Only Ed25519, ECDSA and RSA
from 2048 bits (plus the two security-key forms) are accepted; a key carrying an option such as
command="…" is refused, not cleaned, and every line is written with restrict, so a key
can move files and nothing else. A suspended account's keys are moved aside and come back on
resume, because a locked password alone does not stop a key login.
Hosts installed before this change need the installer re-run to get the line and the directory. Adding keys from the panel's SFTP screen is not in this release yet.
FTPS ships installed and switched off
The installer installs vsftpd, and FTPS works end to end. It is installed off. Nothing listens, no configuration is rendered, and the distribution's own service is masked until an administrator turns FTPS on from the panel.
Turning FTPS on waits up to 10 seconds for vsftpd to answer before it judges the result. If the daemon never answers, the change is rolled back. A first enable leaves no configuration behind, so a second attempt starts clean and the panel never reports success while the daemon is down.
The encryption line says what was measured
The FTPS server status shows an Encryption line read from the daemon's configuration on the host, not from the panel's own setting. It shows one of three states:
- TLS required for logins and transfers: the daemon refuses a login or a transfer that is not encrypted.
- TLS is optional: logins and transfers can travel unencrypted: the daemon accepts both, so a client that does not ask for TLS sends its password in the clear. Fix this before sharing the server.
- Not observed: the configuration could not be read: the agent could not read the daemon's configuration, so the panel says neither. This is not a warning that TLS is off. Check that the agent is running and refresh.
For integrations, GET /api/v1/ftps-server carries this as tlsEnforcement: enforced,
notEnforced or notObserved. It replaces forcedTls, which could only say yes or no and so showed
an unreadable configuration as "no".
We say this plainly because an earlier release note claimed no FTP daemon was installed at all, and an operator who believes that will not look for one on their own host.
Backups
Scheduled account backups, with retention policies, run from the panel. They do not cover a database drop. See databases, where dropping is final and the drop path checks nothing.
When a scheduled backup goes wrong, the administrator gets a mail
Each of these used to fail silently. Each now sends one alert to the address under Outgoing mail → Send alerts to, in English, Russian or Armenian:
- A scheduled run failed: it ended without an archive. The account keeps its previous copies.
- A destination refused: the schedule's destination could not be used, so no run was started.
- A run was missed: a schedule started more than 30 minutes after its named hour, usually because the panel was not running then. A panel that was off for several nights catches up with one run and sends one alert naming the first missed hour.
Each alert is written to the audit journal as "Alert raised", with "Mail skipped: no SMTP settings" beside it when no alert address is configured. Manual backups send no mail, since whoever started one is watching its task. Archives are not yet verified after they are written. Until they are, a damaged archive shows up at restore time, which checks every database dump before loading it.
What nothing re-checks
The installer writes the one Match Group block that turns an SFTP account into a jailed login,
and running it again is safe. Nothing checks it afterwards. If a hand edit or a distribution upgrade
removes the block from sshd_config, every file transfer customer gets a
full shell session, and the panel still reports every login as healthy.
If you edit sshd_config on a Maran host, check that the block is still there.