Mdev 5536 systemd socket activation notification

[grooverdan]

Adds system functionality to mariadb including socket activation and notification.

[janlindstrom]

Please note that buf_dump_status does not use ib_logf currently, instead fprintf directly.

[grooverdan]

@janlindstrom d0619cd too invasive? could boost it with asserts that enum status_severity enumberations match ib_log_level_t

[grooverdan]

Hi, Did you test the output,

no

why you did not remove ut_print_timestamp(stderr); ?

hadn't tested it.

Have tested now - cleaned up as per f7aad0b.

[ghost]

The support-files/mariadb-systemd-start file should be removed as it is a security hazard. The init system should not automatically create the database. This should be left up to the machine's administrator. As the mysql_install_db script opens a password-less root user, this is unsafe. https://bugs.gentoo.org/show_bug.cgi?id=483526 shows a discussion from Gentoo's change of manual systemd integration by our distro.

[grooverdan]

@grknight , thanks so much for looking a this. I've looked at at the gentoo scripts/service files. I quite like the idea of the mysqld_at service and will probably incorporate its concepts into whatever is accepted as the single instance version.

Is a passwordless root the only concern? Its a good point. As an alternative I'm thinking about having unix_socket plugin installed by default and root@localhost via unix_socket as the default grant, passwordless yes, but can only be accessed via the unix root user. Its the way the current debian/ubuntu packages are (though it seems to have some packaging problems - might be my error there).

With socket activation (main benefit I see is more flexible network configuration, easy early runtime dependencies, and systemd restart on process dying), is mariadb-socket-convert sufficient? Should I be trying to script the migration of less used mysqld_safe options to their equivalent (as per mariadb.service)?

[ghost]

On 4/29/2015 2:19 AM, Daniel Black wrote:

@grknight https://github.com/grknight , thanks so much for looking a
this. I've looked at at the gentoo scripts/service files. I quite like
the idea of the mysqld_at service and will probably incorporate its
concepts into whatever is accepted as the single instance version.

Is a passwordless root the only concern? Its a good point. As an
alternative I'm thinking about having unix_socket plugin installed by
default and root@localhost via unix_socket as the default grant,
passwordless yes, but can only be accessed via the unix root user. Its
the way the current debian/ubuntu packages are (though it seems to have
some packaging problems - might be my error there).

Yes that is a concern. unix_socket also can cause a confusion of users
who are used to 'mysql -u root -p' from another user as it will fail
when using unix_socket. This is because the auth plugin system does not
fall back to stored passwords.

With socket activation (main benefit I see is more flexible network
configuration, easy early runtime dependencies, and systemd restart on
process dying), is mariadb-socket-convert sufficient? Should I be trying
to script the migration of less used mysqld_safe options to their
equivalent (as per mariadb.service)?

Systemd's Socket activation is not right for MariaDB IMO.
Think about recovery from a crash, be it system (environment or other)
or segfault. Pending database calls would be waiting or even timeout
unnecessarily since we cannot guarantee near-instant start up and action.

I know that XtraDB/InnoDB can be faster in recent times to recover, but
may not always be on larger systems or if the recovery fails.

Also, socket activation + full SST with galera may not be pretty since
it depends on network transfers.

[grooverdan]

On 22 May 2015 at 00:46, Brian Evans [email protected] wrote:

On 4/29/2015 2:19 AM, Daniel Black wrote:

@grknight https://github.com/grknight , thanks so much for looking a
this. I've looked at at the gentoo scripts/service files. I quite like
the idea of the mysqld_at service and will probably incorporate its
concepts into whatever is accepted as the single instance version.

Is a passwordless root the only concern?....

Yes that is a concern. unix_socket also can cause a confusion of users

This is getting off topic for this PR.

With socket activation (main benefit I see is more flexible network
configuration, easy early runtime dependencies, and systemd restart on
process dying), is mariadb-socket-convert sufficient? Should I be trying
to script the migration of less used mysqld_safe options to their
equivalent (as per mariadb.service)?

Systemd's Socket activation is not right for MariaDB IMO.
Think about recovery from a crash, be it system (environment or other)
or segfault. Pending database calls would be waiting or even timeout
unnecessarily since we cannot guarantee near-instant start up and action.

You may be misunderstand the role that systemd socket activation plays. I
passes the Listening socket FD only to the mysqld process and lets mysqld
to all the accepts.
Because happens, its the same impact with or without socket activation. A
slow start up may have a listening socket there slightly earlier however
there is acceptance until mysqld is ready. While smaller foot print
services use socket activation to trigger the start of a service on connect
I don't expect this to be used with MariaDB. Hence the mariadb.service
definition that contains After=network.target, After=syslog.target. If you
still don't like socket activation, use type=notify. Its here too.

I know that XtraDB/InnoDB can be faster in recent times to recover, but
may not always be on larger systems or if the recovery fails.

Believe me I know, I work on some fairly large systems.

Also, socket activation + full SST with galera may not be pretty since

it depends on network transfers.

As noted in the MDEVs there is more work to do with galera. Its not too
much, just need to map a socket to a config var. Given this patch has
waited over 5 months with only some mediocure review on small parts of the
patch I'm not going to rush to complete the galera component right now.

[vuvova]

I'd prefer three-value option, YES/NO/AUTO, with AUTO being the default.

[grooverdan]

Did the clean ups.

10.0 work stalled at the thought of including systemd notify_* function into mariadb. Will need to purge out the socket activation and enhance mariadb-service-convert to port port/socket options too

attempted to create a 10.1 merge however it seemed to encompase everything from 10.0 that hasn't yet been merged to 10.1 https://github.com/MariaDB/server/compare/10.1...openquery:MDEV-5536-systemd-socket-activation-notification-10.1?expand=1

Where do you met to go from here?

[davidstrauss]

I'd like to leave the code there so distros can move to socket activation if they want and leave the systemd scripts as Type=Notify until 10.1(?) ?

I'm a bit confused at this comment. I don't see why you would stop running the service as Type=notify even if you do use socket activation.

[grooverdan]

I'm a bit confused at this comment. I don't see why you would stop running the service as Type=notify even if you do use socket activation.

Quite right. I though from the comments in on of the MDEV it was ignored when socket activation was used but yes I will keep it the same.

I can't add a dependency on libsystemd-daemon0 in a stable branch, even if compile time configurable(?),

So at the moment I've got to rebase/merge this for 10.1 and/or remove socket activation and include the sd_notify calls from libsystemd-daemon0 in 10.0. And a bit more testing just to be sure - especially on packaging changes.

[vuvova]

I'm sorry, but I don't quite understand the state of this PR. You've created #83 for 10.1, fine. Is this one still for 10.0? Or is it obsolete and should be closed?

[grooverdan]

can be closed.