Using Systemd in DevStack

By default DevStack is run with all the services as systemd unit files. Systemd is now the default init system for nearly every Linux distro, and systemd encodes and solves many of the problems related to poorly running processes.

Why this instead of screen?

The screen model for DevStack was invented when the number of services that a DevStack user was going to run was typically < 10. This made screen hot keys to jump around very easy. However, the landscape has changed (not all services are stoppable in screen as some are under Apache, there are typically at least 20 items)

There is also a common developer workflow of changing code in more than one service, and needing to restart a bunch of services for that to take effect.

Unit Structure

Note

Originally we actually wanted to do this as user units, however there are issues with running this under non interactive shells. For now, we’ll be running as system units. Some user unit code is left in place in case we can switch back later.

All DevStack user units are created as a part of the DevStack slice given the name devstack@$servicename.service. This makes it easy to understand which services are part of the devstack run, and lets us disable / stop them in a single command.

Manipulating Units

Assuming the unit n-cpu to make the examples more clear.

Enable a unit (allows it to be started):

sudo systemctl enable devstack@n-cpu.service

Disable a unit:

sudo systemctl disable devstack@n-cpu.service

Start a unit:

sudo systemctl start devstack@n-cpu.service

Stop a unit:

sudo systemctl stop devstack@n-cpu.service

Restart a unit:

sudo systemctl restart devstack@n-cpu.service

See status of a unit:

sudo systemctl status devstack@n-cpu.service

Operating on more than one unit at a time

Systemd supports wildcarding for unit operations. To restart every service in devstack you can do that following:

sudo systemctl restart devstack@*

Or to see the status of all Nova processes you can do:

sudo systemctl status devstack@n-*

We’ll eventually make the unit names a bit more meaningful so that it’s easier to understand what you are restarting.

Querying Logs

One of the other major things that comes with systemd is journald, a consolidated way to access logs (including querying through structured metadata). This is accessed by the user via journalctl command.

Logs can be accessed through journalctl. journalctl has powerful query facilities. We’ll start with some common options.

Follow logs for a specific service:

sudo journalctl -f --unit devstack@n-cpu.service

Following logs for multiple services simultaneously:

sudo journalctl -f --unit devstack@n-cpu.service --unit devstack@n-cond.service

or you can even do wild cards to follow all the nova services:

sudo journalctl -f --unit devstack@n-*

Use higher precision time stamps:

sudo journalctl -f -o short-precise --unit devstack@n-cpu.service

By default, journalctl strips out “unprintable” characters, including ASCII color codes. To keep the color codes (which can be interpreted by an appropriate terminal/pager - e.g. less, the default):

sudo journalctl -a --unit devstack@n-cpu.service

When outputting to the terminal using the default pager, long lines will be truncated, but horizontal scrolling is supported via the left/right arrow keys. You can override this by setting the SYSTEMD_LESS environment variable to e.g. FRXM.

You can pipe the output to another tool, such as grep. For example, to find a server instance UUID in the nova logs:

sudo journalctl -a --unit devstack@n-* | grep 58391b5c-036f-44d5-bd68-21d3c26349e6

See man 1 journalctl for more.

Debugging

Using pdb

In order to break into a regular pdb session on a systemd-controlled service, you need to invoke the process manually - that is, take it out of systemd’s control.

Discover the command systemd is using to run the service:

systemctl show devstack@n-sch.service -p ExecStart --no-pager

Stop the systemd service:

sudo systemctl stop devstack@n-sch.service

Inject your breakpoint in the source, e.g.:

import pdb; pdb.set_trace()

Invoke the command manually:

/usr/local/bin/nova-scheduler --config-file /etc/nova/nova.conf

Some executables, such as nova-compute, will need to be executed with a particular group. This will be shown in the systemd unit file:

sudo systemctl cat devstack@n-cpu.service | grep Group
Group = libvirt

Use the sg tool to execute the command as this group:

sg libvirt -c '/usr/local/bin/nova-compute --config-file /etc/nova/nova-cpu.conf'

Using remote-pdb

remote-pdb works while the process is under systemd control.

Make sure you have remote-pdb installed:

sudo pip install remote-pdb

Inject your breakpoint in the source, e.g.:

import remote_pdb; remote_pdb.set_trace()

Restart the relevant service:

sudo systemctl restart devstack@n-api.service

The remote-pdb code configures the telnet port when set_trace() is invoked. Do whatever it takes to hit the instrumented code path, and inspect the logs for a message displaying the listening port:

Sep 07 16:36:12 p8-100-neo devstack@n-api.service[772]: RemotePdb session open at 127.0.0.1:46771, waiting for connection ...

Telnet to that port to enter the pdb session:

telnet 127.0.0.1 46771

See the remote-pdb home page for more options.

Known Issues

Be careful about systemd python libraries. There are 3 of them on pypi, and they are all very different. They unfortunately all install into the systemd namespace, which can cause some issues.

  • systemd-python - this is the upstream maintained library, it has a version number like systemd itself (currently 234). This is the one you want.

  • systemd - a python 3 only library, not what you want.

  • python-systemd - another library you don’t want. Installing it on a system will break ansible’s ability to run.

If we were using user units, the [Service] - Group= parameter doesn’t seem to work with user units, even though the documentation says that it should. This means that we will need to do an explicit /usr/bin/sg. This has the downside of making the SYSLOG_IDENTIFIER be sg. We can explicitly set that with SyslogIdentifier=, but it’s really unfortunate that we’re going to need this work around. This is currently not a problem because we’re only using system units.

Future Work

user units

It would be great if we could do services as user units, so that there is a clear separation of code being run as not root, to ensure running as root never accidentally gets baked in as an assumption to services. However, user units interact poorly with devstack-gate and the way that commands are run as users with ansible and su.

Maybe someday we can figure that out.

References