Four People, Four Datacenters, Three Thousand Servers
In 2015 I joined Irideos (then KPNQwest Italia) as a Cloud Architect. The job sounded fancy. The reality was four datacenters, roughly three thousand servers, a team of four engineers, and a budget that could generously be described as “creative.”
This is how we made it work, and what I still carry from that experience today.
Automating 3,000 servers with four people#
With 3,000 servers and 4 people, you can’t afford to do anything by hand. Nobody SSHes into a box or hand-crafts a configuration. You automate everything or you drown.
For us, Infrastructure as Code was a matter of survival.
We rebuilt the entire operational stack from scratch:
- Monitoring: Icinga2, because Nagios was showing its age and we needed something that could scale without a dedicated team to babysit it
- Logging: ELK stack with centralized authentication, because when something breaks at 3 AM, you need to search logs, not read them
- Identity: FreeIPA, because managing credentials across 3,000 servers with spreadsheets is a form of professional self-harm
- Automation: Puppet initially, then Ansible, orchestrated through Satellite and GitLab CI/CD
- DNS: PowerDNS for recursion, Bind for authoritative, including DDoS mitigation that we learned about the hard way
The lesson I keep relearning#
Looking back, the best infrastructure decisions we made came from having no money, not from being smart. When you can’t buy your way out of a problem, you have to actually think about it.
Every tool we chose had to earn its place, and every automation had to save more time than it cost to build. Processes had to be simple enough that any of the four of us could handle them at 3 AM, right after being woken up by a page.
I still catch myself applying this filter: what is the simplest thing that actually works? I do it even at Google Cloud, where the temptation to throw services at a problem is constant and the catalog is infinite.
Then and now#
Each color is one job, in both stacks: pink for CI/CD, amber for automation, blue for monitoring, and green for where the workloads run.
different tools"| NOW classDef ci fill:#3b2b28,stroke:#d3869b,stroke-width:1.5px,color:#eceae5 classDef auto fill:#40361d,stroke:#eec35e,stroke-width:1.5px,color:#eceae5 classDef mon fill:#2d3128,stroke:#83a598,stroke-width:1.5px,color:#eceae5 classDef run fill:#343621,stroke:#a8c373,stroke-width:2px,color:#eceae5 class D1,D2 ci class A1,A2 auto class C1,C2 mon class B1,B2 run style THEN fill:#1f1c14,stroke:#8c8677,stroke-width:1.5px,color:#eceae5 style NOW fill:#1f1c14,stroke:#8c8677,stroke-width:1.5px,color:#eceae5 linkStyle 0,1,2,3,4,5 stroke:#b8b5ad,stroke-width:1.5px linkStyle 6 stroke:#b8b5ad,stroke-width:1.5px,stroke-dasharray:5 4,color:#b8b5ad
What this taught me about cloud#
When I moved to cloud consulting (first Red Hat, then Google), I kept running into the same problem we had at Irideos: how do you keep things running at scale without hiring an army?
The tools are different (Terraform instead of Puppet, GKE instead of bare metal, Cloud Monitoring instead of Icinga2), but the question is the same. And honestly, the companies that struggle most with cloud are the ones that never had to work under real constraints. They have budget, so they buy complexity, and then they need 20 people to operate what 4 people could handle with simpler choices.
This is the first post on this blog. More war stories, technical write-ups, and unsolicited opinions to follow.
Update, November 2026: after my years at Google Cloud I joined Traefik Labs as a Partner Solution Engineer. Same filter, different tools again.