In cases where pgloader is not enough and we need to do some additional
migration work after it, we can now use
`additional_psql_statements_list` and
`additional_psql_statements_db_name`.
This is to be used when migrating `matrix-registration`'s data at the
very least.
We were running into conflicts, because having initialized
the roles (users) and databases, trying to import leads to
errors (role XXX already exists, etc.).
We were previously ignoring the Synapse database (`homeserver`)
when upgrading/importing, because that one gets created by default
whenever the container starts.
For our additional databases, it's a similar situation now.
It's not created by default as soon as Postgres starts with an empty
database, but rather we create it as part of running the playbook.
So we either need to skip those role/database creation statements
while upgrading/importing, or to avoid creating the additional database
and rely on the import for that. I've gone for the former, because
it's already similar to what we were doing and it's simpler
(it lets `setup_postgres.yml` be the same in all scenarios).
Instead of passing the connection string, we can now pass a name of a
variable, which contains a connection string.
Both are supported for having extra flexibility.
Since we'll likely have generic SQLite database importing
via [pgloader](https://pgloader.io/) for migrating bridge
databases from SQLite to Postgres, we'd rather avoid
calling the "import Synapse SQLite database" command
as just `--tags=import-sqlite-db`.
Similarly, for the media store, we'd like to mention that it's
related to Synapse as well.
We'd like to be more explicit, so as to be less confusing,
especially in light of other homeserver implementations
coming in the future.
While these modules are really nice and helpful, we can't use them
for at least 2 reasons:
- for us, Postgres runs in a container on a private Docker network
(`--network=matrix`) without usually being exposed to the host.
These modules execute on the host so they won't be able to reach it.
- these modules require `psycopg2`, so we need to install it before
using it. This might or might not be its own can of worms.
The tasks in `create_additional_databases.yml` will likely
ensure `matrix-postgres.service` is started, etc.
If no additional databases are defined, we'd rather not execute that
file and all these tasks that it may do in the future.
> Invalid data passed to 'loop', it requires a list, got this instead: matrix_postgres_additional_databases. Hint: If you passed a list/dict of just one element, try adding wantlist=True to your lookup invocation or use q/query instead of lookup.
Well, or working around it, as I've done in this commit (which seems
more sane than `wantlist=True` stuff).
To avoid needing to have `jq` installed on the machine, we could:
- try to run jq in a Docker container using some small image providing
that
- better yet, avoid `jq` altogether
Moving it above the "uninstalling" set of tasks is better.
Extracting it out to another file at the same time, for readability,
especially given that it will probably have to become more complex in
the future (potentially installing `jq`, etc.)
If a service is enabled, a database for it is created in postgres with a uniqque password. The service can then use this database for data storage instead of relying on sqlite.
The Docker 19.04 -> 20.10 upgrade contains the following change
in `/usr/lib/systemd/system/docker.service`:
```
-BindsTo=containerd.service
-After=network-online.target firewalld.service containerd.service
+After=network-online.target firewalld.service containerd.service multi-user.target
-Requires=docker.socket
+Requires=docker.socket containerd.service
Wants=network-online.target
```
The `multi-user.target` requirement in `After` seems to be in conflict
with our `WantedBy=multi-user.target` and `After=docker.service` /
`Requires=docker.service` definitions, causing the following error on
startup for all of our systemd services:
> Job matrix-synapse.service/start deleted to break ordering cycle starting with multi-user.target/start
A workaround which appears to work is to add `DefaultDependencies=no`
to all of our services.
`-v` magically creates the source destination as a directory,
if it doesn't exist already. We'd like to avoid this magic
and the potential breakage that it might cause.
We'd rather fail while Docker tries to find things to `--mount`
than have it automatically create directories and fail anyway,
while having contaminated the filesystem.
There's a lot more `-v` instances remaining to be fixed later on.
This is just some start.
Things like `matrix_synapse_container_additional_volumes` and
`matrix_nginx_proxy_container_additional_volumes` were not changed to
use `--mount`, as options for each one are passed differently
(`ro` is `ro`, but `rw` doesn't exist and `slave` is `bind-propagation=slave`).
To avoid breaking people's custom volume mounts, we keep it as it is for now.
A deficiency with `--mount` is that it lacks the `z` option (SELinux
ownership changes), and some of our `-v` instances use that. I'm not
sure how supported SELinux is for us right now, but it might be,
and breaking that would not be a good idea.
If the SQLite database was from an older version of Synapse, it appears
that Synapse would try to run migrations on it first, before importing.
This was failing, because the file wasn't writable.
Hopefully, this fixes the problem.
We recently had a report of the Postgres backup container's log file
growing the size of /var/lib/docker until it ran out of disk space.
Trying to prevent similar problems in the future.
Depending on the distro, common commands like sleep and chown may either
be located in /bin or /usr/bin.
Systemd added path lookup to ExecStart in v239, allowing only the
command name to be put in unit files and not the full path as
historically required. At least Ubuntu 18.04 LTS is however still on
v237 so we should maintain portability for a while longer.
the current version fails the import, because the volume for the media is missing. It still fails if you have the optional shared secret password provider is enabled, so that might need another mount. Commenting out the password provider in the hoimeserver.yaml during the run works as well.
We do use some `:latest` images by default for the following services:
- matrix-dimension
- Goofys (in the matrix-synapse role)
- matrix-bridge-appservice-irc
- matrix-bridge-appservice-discord
- matrix-bridge-mautrix-facebook
- matrix-bridge-mautrix-whatsapp
It's terribly unfortunate that those software projects don't release
anything other than `:latest`, but that's how it is for now.
Updating that software requires that users manually do `docker pull`
on the server. The playbook didn't force-repull images that it already
had.
With this patch, it starts doing so. Any image tagged `:latest` will be
force re-pulled by the playbook every time it's executed.
It should be noted that even though we ask the `docker_image` module to
force-pull, it only reports "changed" when it actually pulls something
new. This is nice, because it lets people know exactly when something
gets updated, as opposed to giving the indication that it's always
updating the images (even though it isn't).
Previously, it only mentioned exposing for psql-usage purposes.
Realistically, it can be used for much more. Especially given that
psql can be easily accessed via our matrix-postgres-cli script,
without exposing the container port.
Reasoning is the same as for matrix-org/synapse#5023.
For us, the journal used to contain `docker` for all services, which
is not very helpful when looking at them all together (`journalctl -f`).
Using `docker_container` with a `cap_drop` argument requires
Ansible >=2.7.
We want to support older versions too (2.4), so we either need to
stop invoking it with `cap_drop` (insecure), or just stop using
the module altogether.
Since it was suffering from other bugs too (not deleting containers
on failure), we've decided to remove `docker_container` usage completely.
We run containers as a non-root user (no effective capabilities).
Still, if a setuid binary is available in a container image, it could
potentially be used to give the user the default capabilities that the
container was started with. For Docker, the default set currently is:
- "CAP_CHOWN"
- "CAP_DAC_OVERRIDE"
- "CAP_FSETID"
- "CAP_FOWNER"
- "CAP_MKNOD"
- "CAP_NET_RAW"
- "CAP_SETGID"
- "CAP_SETUID"
- "CAP_SETFCAP"
- "CAP_SETPCAP"
- "CAP_NET_BIND_SERVICE"
- "CAP_SYS_CHROOT"
- "CAP_KILL"
- "CAP_AUDIT_WRITE"
We'd rather prevent such a potential escalation by dropping ALL
capabilities.
The problem is nicely explained here: https://github.com/projectatomic/atomic-site/issues/203
This makes all containers (except mautrix-telegram and
mautrix-whatsapp), start as a non-root user.
We do this, because we don't trust some of the images.
In any case, we'd rather not trust ALL images and avoid giving
`root` access at all. We can't be sure they would drop privileges
or what they might do before they do it.
Because Postfix doesn't support running as non-root,
it had to be replaced by an Exim mail server.
The matrix-nginx-proxy nginx container image is patched up
(by replacing its main configuration) so that it can work as non-root.
It seems like there's no other good image that we can use and that is up-to-date
(https://hub.docker.com/r/nginxinc/nginx-unprivileged is outdated).
Likewise for riot-web (https://hub.docker.com/r/bubuntux/riot-web/),
we patch it up ourselves when starting (replacing the main nginx
configuration).
Ideally, it would be fixed upstream so we can simplify.
If this is a brand new server and Postgres had never started,
detecting it before we even start it is not possible.
This moves the logic, so that it happens later on, when Postgres
would have had the chance to start and possibly initialize
a new empty database.
Fixes#82 (Github issue)
With this change, the following roles are now only dependent
on the minimal `matrix-base` role:
- `matrix-corporal`
- `matrix-coturn`
- `matrix-mailer`
- `matrix-mxisd`
- `matrix-postgres`
- `matrix-riot-web`
- `matrix-synapse`
The `matrix-nginx-proxy` role still does too much and remains
dependent on the others.
Wiring up the various (now-independent) roles happens
via a glue variables file (`group_vars/matrix-servers`).
It's triggered for all hosts in the `matrix-servers` group.
According to Ansible's rules of priority, we have the following
chain of inclusion/overriding now:
- role defaults (mostly empty or good for independent usage)
- playbook glue variables (`group_vars/matrix-servers`)
- inventory host variables (`inventory/host_vars/matrix.<your-domain>`)
All roles default to enabling their main component
(e.g. `matrix_mxisd_enabled: true`, `matrix_riot_web_enabled: true`).
Reasoning: if a role is included in a playbook (especially separately,
in another playbook), it should "work" by default.
Our playbook disables some of those if they are not generally useful
(e.g. `matrix_corporal_enabled: false`).
As suggested in #63 (Github issue), splitting the
playbook's logic into multiple roles will be beneficial for
maintainability.
This patch realizes this split. Still, some components
affect others, so the roles are not really independent of one
another. For example:
- disabling mxisd (`matrix_mxisd_enabled: false`), causes Synapse
and riot-web to reconfigure themselves with other (public)
Identity servers.
- enabling matrix-corporal (`matrix_corporal_enabled: true`) affects
how reverse-proxying (by `matrix-nginx-proxy`) is done, in order to
put matrix-corporal's gateway server in front of Synapse
We may be able to move away from such dependencies in the future,
at the expense of a more complicated manual configuration, but
it's probably not worth sacrificing the convenience we have now.
As part of this work, the way we do "start components" has been
redone now to use a loop, as suggested in #65 (Github issue).
This should make restarting faster and more reliable.
