All projects04 / 08
One command-centre dashboard with a row per project.
01 / Problem
Several projects share one machine, and I wanted one screen that says whether all of it is fine, including a display that runs unattended on the hypervisor. I also wanted the same view to show what each project is doing, not only CPU and RAM.
02 / How it works
Prometheus scrapes the containers, the VM and the Proxmox host. I wrote two exporters in plain Python: one turns the AnunțulCasei crawler status into metrics, the other reports traffic and errors for the public ONE·WAY demo. Alloy ships container logs to Loki, and Grafana loads dashboards and data sources from JSON files in git.
03 / The hard part
One cAdvisor metric group made the disk read hundreds of MB per second and pushed the load above 20 under its memory cap, so I switched it off after checking that no panel used it. Container panels filter on the Compose project label, so a new stack appears on its own and ad-hoc containers stay off. When the end-of-life log shipper was replaced by Alloy I kept the same labels, so no dashboard query had to change.
04 / Result
The command centre has 63 panels in 8 rows, from the VM and Proxmox host to each project. A second dashboard for AnunțulCasei has a table that states in words what looks off, for example overdue or blocked crawlers, and it compares counts rather than percentages while the numbers are small. Dashboard edits go into git, because changes made only in the UI would sit in a volume without a backup.
How it works
Standard collectors cover the machines. Two small exporters cover what only a project knows about itself.
The same flow as a list, in order.
Stack, by layer
Live from PrometheusOffline