Currently, when starting Mattermost programmatically, it's hard to tell
when the server is actually ready to receive network connections.
This isn't convenient for monitoring (the systemd service status is
"running" although the server is still booting), nor for programatic use
(where a script would need to know when the server is ready to perform
further actions).
To improve this, systemd allow processes to tell when they started
successfully. The launcher waits for this notification before
reporting the service as successfully launched.
The way processes notify systemd is by sending a `READY=1` string over
a standard unix socket, whose path is provided in an environment var.
The systemd service is then told to expect this notification:
```diff
[Service]
-Type=simple
+Type=notify
ExecStart=/home/vagrant/go/bin/platform
```
Now, when starting the server, systemd will actually wait for the server to
be ready before returning the control to the shell.
Additionally, during this time, querying the server status with
`service mattermost status` will report the service as "activating" – before
transitioning to "running" when the server is ready.
When running server tests, the server will exit while some jobs spawned
through a goroutine are still running. This may crash the test harness,
as the jobs try to access a shut down app instance.
Fortunately the fix is easy: we just have to use the same App
goroutine-counting facility than the rest of the method's goroutines.
* Cherry-picking 7b2861de3a to master because enterprise/master has code in it that enterprise/release-4.7 does not, and I don't want enterprise/master to break the next time somebody merges enterprise/release-4.7 to enterprise/master
* Renamed file to match existing scheme
(cherry picked from commit 8c22c5c6c6)
* Cleanup app state on initialization error
When returning an initialization error, the app state was not cleaned
up. This is especially visible during tests, as `appCount` is not
decremented, and makes the new app initialization fail.
* Test the `platform server` command
As the `platform server` command only exits when interrupted by
a signal, it is not possible to test it as the other cobra
commands. Instead we directly test the actual command function.
The internal command handler is slighly refactored to take
a channel in argument, and registers it as the signal handler.
Nothing very different—except than controlling this channel
from the outside allows the test to send the system signal
itself, thus preventing the server to run forever.
* [PTL-7537] implement feature and test
* [PTL-7537] Update feature to post the the room requiring a username flag to be used
* [PTL-7537] update tests with username
* update test to remove changes to the test helper struct
* use the basic team and user
* Added username to ChannelMemberHistory struct in anticipation of supporting GlobalRelay in Compliance Export
* Removed translation from debug output - this makes it complicated to use utils functions from tests in the enterprise repo
* Added an advanced email function that allows for greater control over message details. Updated MessageExport config to support GlobalRelay. Added attachment support to InBucket unit tests
* Moving templates in from enterprise to solve test issues
* Added export format to diagnostics
* Changed email attachment code to use FileBackend so that S3 storage is properly supported
Only a handful of critical errors are present in the codebase.
They all occur during server startup (in `app.StartServer()`).
Currently, when one of these critical error occurs, it is simpled
mentionned in the logs – then the error is discarded, and the app
attempts to continue the execution (and probably fails pretty quickly in
a weird way).
Rather than continuing operations in an unknow state, these errors should
trigger a clean exit.
This commit rewrites critical startup errors to be correctly
propagated, logged, and then terminate the command execution.
Additionnaly, it makes the server return a proper error code to the
shell.
