Day-to-day is 100% automated. But initial deployment - not so much.
digital-seat-198720 is the current name of the GCP project. When switching
projects, you gotta grep through this repository and replace all occurrences by
hand, unfortunately. You gotta create the project manually too, in GCP.
Terraform is used for provisioning. For GCP access, it'll look at your default
credentials, and for Cloudflare - for a file called terraform.tfvars, with
the following contents:
cloudflare_email = "EMAIL"
cloudflare_token = "TOKEN"Look at Terraform documentation for further info.
Just in case, us-east1 is the only fitting "Always Free" region.
Watch out for Terraform provider version pinning, it might result in odd things down the line. Make sure to update regularly.
Terraform's state is stored in a bucket managed by Terraform itself. This is a
bit of a chicken'n'egg, but it is easily addressed. First, comment out the
terraform section at the top of terraform.tf. Then, run terraform init
and terraform apply. Afterwards, uncomment that section and run terraform init again. That's it.
Once provisioned, everything is ready (including DNS), except for the data
disk. It needs to be formatted to ext4. That is a bit of an ordeal on COOS
(which you need to SSH into), but can be done. Just SSH with gcloud compute ssh user@main and run sudo mkfs.ext4 /dev/disk/by-id/google-data.
After that, you also need to supply the secret stuff, by the means of creating
a local file called .secrets of the following shape:
SENDGRID=TOKEN
SLACK=https://hooks.slack.com/services/TOKEN
SENTRY=https://TOKEN@app.getsentry.com/TOKEN
LOADER=loaderio-TOKEN
BF4_ADDRESS=IP:PORT
BF4_PASSWORD=PASSWORDAnd then uploading it by calling scripts/secrets u. This isn't the perfect
solution (something like Vault would be amazing), but a pretty good bang-for-
buck simplicity- and security-wise. FYI, there's a symmetric command,
scripts/secrets d, which pulls the .secrets from the bucket and overwrites
your local copy with it.
Something that Terraform can't do yet is setup a build trigger (GCP's CI). You
need to do that through the GUI or gcloud. Just create one that points at
this repository, and gets triggered by every commit on master, and uses
cloudbuild.yaml for configuration. That's it.
Another quick gotcha (that Terraform may or may not be able to take care of,
it's a bit complicated) is that PROJECT_ID@cloudbuild.gserviceaccount.com
service account needs to be given two extra roles: Compute Instance Admin (beta) and Service Account User. This is needed so that scripts/deploy can
SSH into an instance in the context of a build job. Done through IAM.
And trigger a build manually. This will go through the whole process of building all images and deploying.
The instance self-recovers when rebooted.
At this point, the app is ready, but the database is empty. To import
production data run scripts/db-replace path/to/your/dump.sql.
And voila!
The only thing that I haven't automated is recurring data disk snapshotting. One reason is that it's a whole lot of bother for no gain (nobody is using this website right now), and another reason is that the data does evolve little by little (due to the worker updating BF4 usernames and such), so the snapshots would keep accumulating, potentially into several dollars per month, despite them being incremental.
However, it's super easy to do manually. Just run gcloud compute disks snapshot data, when needed (e.g. before deploying a new migration).
Something to look out for is log accumulation. Currently, I'm not sure how/if Compose handles rotation, and the logs don't go anywhere (everything sensible is paid), so the instance can run out of space in some distant future.
Previously, there was an HTTP-to-HTTPS rule in Cloudflare. But there's no need in it with the new setup, since the LE helper takes care of it (it redirects to HTTPS immediately, with an HSTS header, both the client and the API).
I wanted to take it a step further and outright refuse HTTP on the API, but it isn't necessary, and, more importantly, causes an unresolvable issue with LE. More info.