Grafana Loki explicitly eschews built-in authentication. In fact, its
[documentation][0] states:
> Operators are expected to run an authenticating reverse proxy in front
> of your services.
While I don't really want to require authentication for agents sending
logs, I definitely want to restrict querying and viewing logs to trusted
users.
There are _many_ reverse proxy servers available, and normally I would
choose _nginx_. In this case, though, I decided to try Caddy, mostly
because of its built-in ACME support. I wasn't really happy with how
the `fetchcert` system turned out, particularly using the Kubernetes API
token for authentication. Since the token will eventually expire, it
will require manual intervention to renew, thus mostly defeating the
purpose of having an auto-renewing certificate. So instead of using
_cert-manager_ to issue the certificate and store it in Kubernetes, and
then having `fetchcert` download it via the Kubernetes API, I set up
_step-ca_ to handle issuing the certificate directly to the server. When
Caddy starts up, it contacts _step-ca_ via ACME and handles the
challenge verification automatically. Further, it will automatically
renew the certificate as necessary, again using ACME.
I didn't spend a lot of time optimizing the Caddy configuration, so
there's some duplication there (i.e. the multiple `reverse_proxy`
statements), but the configuration works as desired. Clients may
provide a certificate, which will be verified against the trusted issuer
CA. If the certificate is valid, the client may access any Loki
resource. Clients that do not provide a certificate can only access the
ingestion path, as well as the "ready" and "metrics" resources.
[0]: https://grafana.com/docs/loki/latest/operations/authentication/
The Promtail container image is pretty big, so it takes quite some time
to pull on a slow machine like a Raspberry Pi. Let's increase the
startup timeout so the service is less likely to fail while the image is
still being pulled.
All the stand-alone FCOS hosts now have Promtail running, forwarding
_systemd_ journal messages to Grafana Loki. The Kubernetes nodes will
have Promtail deployed as a Kubernetes pod.
I would really like to come up with a way to define variables for groups
of hosts, so that I do not have to add `promtail: prod.#promtail` to
every host's values file individually...
[Promtail][0] is the log collection agent for Grafana Loki. It reads
logs from various locations, including local files and the _systemd_
journal and sends them to Loki via HTTP.
Loki configuration is a highly-structured YAML document. Thus, instead
of using Tera template syntax for loops, conditionals, etc., we can use
the full power of CUE to construct the configuration. Using the
`Marshal` function from the built-in `encoding/yaml` package, we
serialize the final configuration structure as a string and write it
verbatim to the configuration file.
I have modeled most of the Promtail configuration schema in the
`du5t1n.me/cfg/app/promtail/schema` package. Having the schema modeled
will ensure the generated configuration is valid during development
(i.e. `cue export` will fail if it is not), which will save time pushing
changes to machines and having Loki complain.
The `#promtail` "function" in `du5t1n.me/cfg/env/prod` makes it easy to
build our desired configuration. It accepts an optional `#scrape`
field, which can be used to provide specific log scraping definitions.
If it is unspecified, the default configuration is to scrape the systemd
journal. Hosts with additional needs can supply their own list,
probably including the `promtail.scrape.journal` object in it to get the
default journal scrape job.
[0]: https://grafana.com/docs/loki/latest/send-data/promtail/
According to the [Grafana Loki documentation][0], sending SIGHUP to the
Loki process will instruct it to reload its configuration. This is
necessary in order for it to re-read its server certificate after it has
been renewed.
[0]: https://grafana.com/docs/loki/latest/configure/#reload-at-runtime
Before going into production with Grafana Loki, I want to set it up to
use TLS. To that end, I have configured _cert-manager_ to issue it a
certificate, signed by _DCH CA_. In order to use said certificate,
we need to configure `fetchcert` to run on the Loki server.
The `fetchcert` tool is a short shell script that fetches an X.509
certificate and corresponding private key from a Kubernetes Secret,
using the Kubernetes API. I originally wrote it for the Frigate server
so it could fetch the _pyrocufflink.blue_ wildcard certificate, which is
managed by _cert-manager_. Since then, I have adapted it to be more
generic, so it will be useful to fetch the _loki.pyrocufflink.blue_
certificate for Grafana Loki.
Although the script is rather simple, it does have several required
configuration parameters. It needs to know the URL of the Kubernetes
API server and have the certificate for the CA that signs the server
certificate, as well as an authorization token. It also needs to know
the namespace and name of the Secret from which it will fetch the
certificate and private key. Finally, needs to know the paths to the
files where the fetched data will be written.
Generally, after certificates are updated, some action needs to be
performed in order to make use of them. This typically involves
restarting or reloading a daemon. Since the `fetchcert` tool runs in
a container, it can't directly perform those actions, so it simply
indicates via a special exit code that the certificate has been updated
and some further action may be needed. The
`/etc/fetchcert/postupdate.sh` script is executed by _systemd_ after
`fetchcert` finishes. If the `EXIT_STATUS` environment variable (which
is set by _systemd_ to the return code of the main service process)
matches the expected code, the configured post-update actions will be
executed.
Since *systemd* starts the *reload-udev-rules.service* unit as soon as
any file in the `/run/containers/udev-rules` directory changes, the `cp`
command may start before all of the files have been copied out of the
container. If this happens, some of the rules will not get copied to
the final path, and thus will not be processed by *udev*.
Togive the container a chance to finish copying all of the files before
we process them, we need a bit of a delay. Obviously, this is not a
perfect solution, as it could potentially take longer than 250ms to copy
the files in some cases, but hopefully those cases are rare enough to
not worry about.
In order to simplify the process of adding new template render
instructions to all hosts, I've created a list of templates in the
`env/prod` module. This way, I only have to add templates there, and
all hosts that "inherit" from it will automatically get them.
I do not like how Fedora CoreOS configures `sudo` to allow the *core*
user to run privileged processes without authentication. Rather than
assign the user a password, which would then have to be stored
somewhere, we'll install *pam_ssh_agent_auth* and configure `sudo` to
use it for authentication. This way, only users with the private key
corresponding to one of the configured public keys can run `sudo`.
Naturally, *pam_ssh_agent_auth* has to be installed on the host system.
We achieve this by executing `rpm-ostree` via `nsenter` to escape the
container. Once it is installed, we configure the PAM stack for
`sudo` to use it and populate the authorized keys database. We also
need to configure `sudo` to keep the `SSH_AUTH_SOCK` environment
variable, so *pam_ssh_agent_auth* knows where to look for the private
keys. Finally, we disable the default NOPASSWD rule for `sudo`, if
and only if the new configuration was installed.
Unfortunately, the automatic transfer switch does not seem to work
correctly. When the standby source is a UPS running on battery, it does
*not* switch sources if the primary fails. In other words, when the
power is out and both UPS are running on battery, when the first one
dies, it will NOT switch to the second one. It has no trouble switching
when the second source is mains power, though, which is very strange.
I have tried messing with all the settings including nominal input
voltage, sensitivity, and frequency tolerence, but none seem to have any
effect.
Since it is more important for the machines to shut down safely than it
is to have an extra 10-15 minutes of runtime during an outage, the best
solution for now is to configure the hosts to shut down as soon as the
first UPS battery gets low. This is largely a waste of the second UPS,
but at least it will help prevent data loss.
The `upsrw` command, which is used to set individual UPS configuration
parameters like low battery level, etc., needs a username and password
to authenticate to `upsd`.
Setting `AutoUpdate=registry` will tell Podman to automatically fetch
an updated container image from its corresponding registry and restart
the container. The `podman-auto-update.timer` systemd unit needs to be
active for this to happen on a schedule.
Since the "primary" `upsmon` is always (for our purposes) running on the
same host as `upsd`, there's no reason to specify both values.
All systems need a shutdown command; one is not set by default.
The primary system is the only one that should send notifications.
`dest` is not a valid option for the `--mount` argument to `podman`. To
specify where the target path, only `target`, `destination`, and `dst`
are valid.
`upsmon` is the component of NUT that tracks the status of UPSs and
reacts to their changing by sending notifications and/or shutting down
the system. It is a networked application that can run on any system;
it can run on a different system than `upsd`, and indeed can run on
multiple systems simultaneously.
Each system that runs `upsmon` will need a username and password for
each UPS it will monitor. Using the CUE [function pattern][0], I've
made it pretty simple to declare the necessary values under
`nut.monitor`.
[0]: https://cuetorials.com/patterns/functions/
*collectd* logs to syslog, so its output is lost when it's running in a
container. We can capture messages from it by mounting the journald
syslog socket into the container.
The `/run/udev/rules.d` directory may not always exist, especially at
boot. We need to ensure that it does before we try to copy rules
exported by containers into it, or the unit will fail.
Even with *collectd* configured to report filesystem usage by device, it
still only reports filesystems that are mounted (in its namespace).
Thus, in order for it to report filesystems like `/boot`, these need to
be mounted in the container.
I keep going back-and-forth on whether or not collectd should run in a
container on Fedora CoreOS machines. On the one hand, running it
directly on the host allows it to monitor filesystem usage by mount
point, which is consistent with how non-FCOS machines are monitored.
On the other hand, installing packages on FCOS with `rpm-ostree` is a
nightmare. It's _incredibly_ slow. There's also occasionally issues
installing packages if the base layer has not been updated in a while
and the new packages require an existing package to be updated.
For the NUT server specifically, I have changed my mind again: the
*collectd-nut* package depends on *nut-client*, which in turn depends on
Python. I definitely want to avoid installing Python on the host, but I
do not want to lose the ability to monitor the UPSs via collectd. Using
a container, I can strip out the unnecessary bits of *nut-client* and
avoid installing Python at all. I think that's worth having to monitor
filesystem usage by device instead of by mount point.
Without the `...` prefix, CUE interprets a type enclosed in square
brackets as a list of exactly one of that type. The ellipsis changes it
to mean a list of any number of that type.
I don't want Jenkins to build a new runtime container every time I make
a change to the configuration policy. As such, I've moved the container
image definition and corresponding CI pipeline script to their own
repository.
A bunch of stuff that wasn't schema definitions ended up in the `schema`
package. Rather than split values up in a bunch of top-level packages,
I think it would be better to have a package-per-app model.
Although KCL is unquestionably a more powerful language, and maps more
closely to my mental model of how host/environment/application
configuration is defined, the fact that it doesn't work on ARM (issue
982]) makes it a non-starter. It's also quite slow (owing to how it
compiles a program to evaluate the code) and cumbersome to distribute.
Fortunately, `tmpl` doesn't care how the values it uses were computed,
so we freely change configuration languages, so long as whatever we use
generates JSON/YAML.
CUE is probably a lot more popular than KCL, and is quite a bit simpler.
It's more restrictive (values cannot be overridden once defined), but
still expressive enough for what I am trying to do (so far).
`tmpl` takes a long time to compile on a Raspberry Pi, so I've created a
CI pipeline to build it separately.
`kcl` seems to have a [bug][0] that causes it to include the x86_64
builds of `kclvm_cli` and `libkclvm_cli_cdylib.so` on aarch64. This
naturally doesn't work, so we need to fetch the correct builds
ourselves.
[0]: https://github.com/kcl-lang/cli/issues/31
The only privilege NUT needs is access to the USB device nodes. Using a
device CGroup rule to allow this is significantly better than disabling
all restrictions. Especially since I discovered that `--privileged`
implies `--security-opt label=disable`, effectively disabling SELinux
confinement of the container.
NUT needs some udev rules in order to set the proper permissions on USB
etc. devices so it can run as an otherwise unprivileged user. Since
udev rules can only be processed on the host, these rules need to be
copied out of the container and evaluated before the NUT server starts.
To enable this, the *nut-server* container image copies the rules it
contains to `/etc/udev/rules.d` if that directory is a mount point. By
bind mounting a directory on the host at that path, we can get a copy of
the rules files outside the container. Then, using a systemd path unit,
we can tell the udev daemon to reload and reevaluate its rules.
SELinux prevents processes in containers from writing to
`/etc/udev/rules.d` directly, so we have to use an intermediate location
and then copy the rules files to their final destination.
Need to run `systemctl daemon-reload` after creating or modifying the
`nut-server.container` unit file, so that the corresponding service unit
will be generated.
When `tmpl` runs `systemd-sysusers` after generating the `sysusers.d`
file for NUT, the `/etc/passwd` and `/etc/group` files on the host are
created anew and replaced, which "breaks" the bind mount. Since new
files are put in their place, the container and the host no longer see
the same files. We can work around this by using a symbolic link for
each file, pointing to the respective file in the `/host` directory
(which is the host's `/` directory bind mounted into the container's
namespace). Since the symlinks follow the file by name rather than
inode, the container's view is always in sync with the host's.
Dustin