Dang near every jail on my #freebsd server needs either pgsql or mariadb.
I could stop running databases everywhere by building a database jail, eliminating many small points of small failures by implementing a single point of complete failure.
Which is better? All I know is whatever choice wins, I will lose. #sysadmin
@mwl In my mind it boils down primarily to how much RAM the server has.
If running multiple instances of mariadb/pgsql would put a strain on RAM then there are definite wins to having a single shared instance of each.
If RAM contention is *not* an issue, then you're weighing the security of independent separated instances vs administrative upgrading/configuration/tuning hassle. And only you know where that line falls.
If you use some sort of declarative configuration-management, then the admin headache might be lessened because you can just specify "ZFS ashift=12 on all those ZFS datasets, upgrade all my pgsql version X to version X+1, and ensure they're configured for $BLOCKSIZE pages" and let it babysit all the cattle rather than hand feeding all the pets. Or you can propagate your security foibles at scale. Whatevs…
@gumnos „weighting security“ of independent db instances: what do you base this on?
How are multiple instances of psql in their separate jails more secure than a single psql jail and setup properly according to long established BCP, using separate „databases“ (in psql parlance), proper roles, privileges and so on?
Aside security I can see *some* reason to do that in special cases, for example like training setups, documentation and or writing books about to mention @mwl ‘s specialty.
To compromise another database in a shared Postgres instance, you need to find a Postgres bug or misconfiguration, or you need a jail escape that lets you directly poke at the other jail.
To compromise another database in another jail, you need to find a jail escape.
Basically, any vulnerability that lets you attack an isolated Postgres instance in another jail will also let you compromise a shared Postgres instance in another jail, but the converse is not true.
