Here's a number that changes how you read every hosting dashboard: a month is 730 hours, not 30 days. Run the math and "99% uptime" — the number people celebrate — is 7.3 hours of your app being down every single month. On a free tier, nobody guarantees even that.
We ran a small ops practice helping micro-teams (a few hundred dollars of monthly revenue riding on their app) and the free-tier outages followed the same three patterns every time:
1. The hibernation trap. Free tiers sleep idle instances aggressively. Your app "works when you test it" because your tests wake it up. A real customer at 3am hits a 30-second cold start, gives up, and you never see the failed session in your logs — it never got far enough to log anything.
2. The quota cliff. Free-tier bandwidth/execution quotas don't degrade gracefully — they cut off. You discover the cliff from your customers, not your dashboards, because the quota meter is the provider's dashboard, not yours.
3. The no-SLA blind spot. Without an SLA, there's no baseline. "Instance asleep" and "instance dead" are indistinguishable from inside the box. The fix is embarrassingly cheap: one external synthetic check (a cron from a different machine hitting /healthz and logging latency) tells you the difference within a minute.
Two watchdog patterns that paid for themselves immediately:
- Restart-and-report: restart on the known patterns (OOM, cold-start timeout, quota throttle), then post what it did and why. A 3am page becomes a morning notification with an audit trail.
- Outage math in the runbook: for every component, write down "at 99%, this is 7.3h/month down; at 99.9%, 43 minutes; at 99.99%, 4.3 minutes." Teams argue passionately about an extra nine without ever writing down what each nine costs them in dollars or lost users.
The upgrade path that actually matters isn't chasing more free hours — it's shortening time-to-know. A free tier that pages you in 60 seconds beats a paid tier you find out about from a customer.
If you want the checklist version — the external-check, restart-and-report, and outage-math templates we use, plus the 5-minute audit that finds which of your automations would silently die tonight — we packaged the whole ops kit here: Agent Ops 24/7 on Gumroad. Free samples on the storefront too.
What's your record for "the app was down and nobody noticed"? Ours was 40 minutes, and that number is exactly why we built the watchdog.
Top comments (0)