Highlights
- MySQL and MariaDB databases move to a newer release from the database page (MySQL 8.0 → 8.4 → 9.7, MariaDB 10.6+ → 12.3): the data directory is copied, upgraded in place and compared before the switch, with rollback until you finalize.
- A database created without a version now keeps the release it runs when a later NineDeploy release moves the engine's default; existing ones are pinned at start.
- Restoring a backup of a MySQL 9 database works again, and database passwords no longer appear in process lists while the panel or a node backs up, restores, imports or measures a database.
Action needed: No. Update each node agent to this release so that database passwords stay out of its process list too.
Added
- Major version upgrades for MySQL and MariaDB databases. Settings → Major version upgrade,
ninedeploy databases upgrade start <dbId> --to 8.4, orPOST /v1/databases/:id/upgradeswith{ "toVersion": "8.4" }. MySQL goes to 8.4 or 9.7 from 8.0 or later; MariaDB to 10.11, 11.4, 11.8 or 12.3 from 10.6 or later. See DATABASES_BACKUPS.md §12.- They move by copy, not by dump: a dump of every database carries the
mysqlsystem schema, which neither project supports loading into another series. The database stops; a networkless container of its own image takes the safety dump and the fingerprint, shuts down slowly and must exit cleanly; its data directory is copied into a new volume, and the new release upgrades the copy in place on start. The old volume is never written to, so a rollback is possible until you finalize. Users, grants and every database come along. - MySQL never skips an LTS series: 8.0 → 9.7 starts 8.4 on the copy first.
- Before the switch the copy is compared with the original: the exact row count of every table, the number of indexes, constraints, views, triggers, routines, events and databases, and every account. Root must log in over TCP with the stored password.
- An upgrade to MySQL 8.4 or later is refused, before anything stops, while an account still uses
mysql_native_password(8.4 disables it, 9.0 removed it). The refusal names the accounts. - Checked on a real panel: MySQL 8.0 → 9.7 (through 8.4) in 35 s with identical rows, the trigger, procedure, view, event and AUTO_INCREMENT working and an app account logging in over the network; rollback to 8.0; 8.0 → 8.4 in 25 s; MariaDB 10.11 → 11.8 in 20 s with identical rows and its accounts.
- The API adds
fromSeries/toSeriesto an upgrade, andengine,currentSeriesandversionsto the options;toMajorstill works for PostgreSQL.
- They move by copy, not by dump: a dump of every database carries the
Fixed
The PostgreSQL credential check accepted any password. After a dump import, and before a recovered (PITR) database or a preview copy is reported ready, the panel checks that it can still sign in with the password it holds. For PostgreSQL it asked over 127.0.0.1, which the official image trusts like the local socket, so the check passed whatever the password was, and an import that changed the
ninerole's password endedcompletedinstead of warning. It now signs in over the container's own network address, where the password is required, read from the container's own/etc/hostsrather than Docker's DNS (another container on a shared network could answer to the database's name), and it refuses a server that asks for the password in clear text (PGREQUIREAUTH=scram-sha-256,md5). A changed password ends the importcompleted_with_warnings. Checked live in both directions, and against PostgreSQL 14, 16, 17 (Alpine), 18 and pgvector.Restoring a backup of a MySQL 9 database failed. The restore ran
mysql -e "source <file>", and MySQL 9's client no longer runs client commands such assourcein-e(it answers a syntax error) unless given--commands, which older clients refuse. MySQL's default is 9.7, so a backup of a database created with the default could not be restored from the panel or on a node. The dump now goes to the client on stdin, as dump imports already did. A second fault hid behind the first: MySQL 9's images run with GTIDs on, somysqldumpwroteSET @@GLOBAL.GTID_PURGED=…into every backup, and restoring it into the server that wrote it fails ("ERROR 3546 … must not overlap with @@GLOBAL.GTID_EXECUTED"). Backups are now taken with--set-gtid-purged=OFF, and a restore drops that statement from its temporary copy first, so backups taken before this release restore too. MariaDB and MySQL 8 were not affected.Database passwords no longer appear in process lists. The size query, the credential check, the backup, the restore and the dump import of every engine that has a password put it in the command line (
--password=…,-p …,-a …,-e PGPASSWORD=…), which any local user on the host can read for as long as the command runs (ps,/proc/<pid>/cmdline), on the panel host and on a node. The password now travels in the environment of thedockercommand, whichdocker exec -e NAMEforwards by name: PostgreSQL readsPGPASSWORD, the Redis familyREDISCLI_AUTHand ClickHouseCLICKHOUSE_PASSWORD; the MySQL and MariaDB clients read a[client]option file andmongodump/mongorestorea YAML config, both on file descriptor 3, so stdin stays free for a dump;mongoshsigns in from a script that reads the password from its environment. Checked against MySQL 8.0 and 9.7, MariaDB 10.6 and 12.3, PostgreSQL 18, MongoDB 7.0 and 8.0, Redis 8.8, Valkey 9.1, KeyDB 6.3.4, Dragonfly 1.34.2 and ClickHouse 23.8, 24.8 and 25.8, with a password holding#"\;$'and spaces, and with a wrong password refused. On a node this takes the new agent. Not yet covered: the Redis family's server still receives its password as--requirepassin its container command, which Redis hides from its process title but Valkey (through its entrypoint) and Dragonfly show.A database without a version no longer follows the engine's default across releases. Such a database ran whatever default the panel had when its container was next recreated (a restart after a crash, a limit change, a restore), so a release that moved the default would start, for example, MySQL 9.7 on 8.4 data and upgrade it in place with no way back, or start a newer Redis that writes files the old one cannot read. New databases (from the API, the CLI, the Hub and templates) are created with the default tag. At start, the panel gives every existing database without a version the tag of the image its container runs. A stopped database (Stop removes the container) gets today's default only when its volume does not exist or was created by today's default image, and a node database gets today's default; Docker not answering, an image no tag reproduces, or a stopped database whose volume an older default created leaves it as it was, logged, and tried again at the next start. A
reuseExistingretry without a version resumes the pinned row. The engine table in DATABASES_BACKUPS.md §1 now shows the real defaults (PostgreSQL 18, MySQL 9.7, MariaDB 12.3, Redis 8.8, Valkey 9.1).