Running a Hosting Business: Hard Lessons from the Trenches

Initial Setup Choices

Two years ago the first production rack went live with a handful of Linode instances paired to bare-metal servers rented from Hetzner. The cloud layer handled quick customer onboarding but collapsed under sustained database load during peak hours. Moving the primary PostgreSQL tier to dedicated Hetzner AX41 boxes eliminated noisy neighbors and cut the monthly bill by roughly forty percent. Network diagrams were redrawn on paper before any cables were run, and every new server received a standard Debian install scripted through a simple shell bootstrap. That early discipline kept later migrations from turning into all-night marathons.

Handling Unexpected Downtime

Last quarter an uplink flap at the primary facility knocked three racks offline for ninety minutes. No automated failover existed at the time, so customers received only a vague status page update. The fix involved standing up a second Proxmox cluster in a separate data center with live migration scripts triggered by simple health checks. Recovery time on the next incident dropped below ten minutes. We also added daily rsync snapshots to an off-site box and tested restores every month. The process exposed how many customer containers lacked proper volume labels, a detail fixed before the next billing cycle.

Monitoring Gaps That Surfaced

Prometheus was added after the outage to track disk I/O and network saturation across both sites. Alerts now fire to a dedicated PagerDuty rotation instead of email inboxes that get ignored overnight. The same stack later caught a failing drive on a storage node before it took down any virtual machines.

Optimizing for Cost Efficiency

Early customers received oversized virtual machines with generous CPU and RAM allocations that stayed idle most of the month. Prometheus graphs later showed seventy percent of instances running below twenty-five percent utilization. Right-sizing those workloads and shifting steady traffic onto reserved Hetzner dedicated servers freed budget for an extra cold-standby node each quarter. Power and bandwidth invoices dropped noticeably once we stopped paying for unused cloud cores. Reserved capacity contracts with Hetzner require twelve-month commitments, but the savings compound quickly when you track actual usage rather than peak projections.

Security Practices That Actually Worked

Default firewall rules on customer instances were routinely ignored, leading to repeated SSH brute-force attempts logged in the thousands. An Ansible playbook now enforces key-only authentication, disables password login, and opens only ports 80 and 443 on every new virtual machine. After rollout, Fail2Ban entries fell by ninety-five percent inside the first month. Separate playbooks handle unattended security updates for the host layer while customer containers receive a weekly report of packages that need attention. The same automation also rotates API tokens for the billing system every ninety days without manual intervention.

Scaling Without Breaking Everything

At five hundred active customers, manual Nginx configuration edits on each reverse proxy became unsustainable. A single misplaced semicolon once took the entire frontend offline during a routine change. Moving every config file into a Git repository with automated deploys through Ansible Tower removed that failure mode. New nodes now join the load-balancer pool after a single command and inherit the exact same rule set. The repository also serves as documentation; anyone can see the history of rate-limit adjustments made six months earlier when a customer experienced a traffic spike from a marketing campaign.

Client Expectations Management

Support tickets frequently demanded uptime guarantees that shared hardware could never meet. Clear SLAs were introduced that tie response times to the service tier purchased, with shared instances carrying only best-effort language. Customers who needed stronger commitments moved to dedicated servers at a predictable price increase. The change reduced ticket volume by a third and improved satisfaction scores because promises finally matched reality. Quarterly reviews now include simple graphs of actual versus promised availability so both sides stay aligned without surprise invoices.

This is Allan Ali for Sylt.ing.