All articles
Infra4 min read

What keeps an app standing: backups, alerts, plan B

Shipping is 30% of the job. The rest: knowing a site is down before your customers do, being able to roll back, surviving the loss of a server, and respecting people's data.

by Vidhal Elame
#Backups#Monitoring#GDPR

Under the same viral post I mentioned in part 4, someone listed everything people forget when they boast about putting a server online in fifteen minutes: monitoring, redundancy, backups, security, networking, GDPR… The tone was annoyed, but the substance was right: shipping is only part of the job. Running it is another.

I went through that list point by point. Here's what I set up, and why.

Backups: the ones you actually tested

My databases live with a managed Postgres provider. On its free plan, it lets you go back 6 hours. A deletion made by mistake in the evening and noticed the next morning would be lost.

So every night the server exports the three databases, compresses them and sends them to object storage, off the server. One folder a day, 30 days of history, and older ones delete themselves. If a backup fails, I get an e-mail.

Database ADatabase BDatabase CEvery nightdump + compressObject storageone folder a dayRestore testidentical rows30 days kept, older ones deleted
Each night the three databases are exported, compressed and sent off the server. A backup you never restored is a hope, not a backup — so the restore was tested.

That's not the important part. A backup you never restored isn't a backup, it's a hope. So I tested it for real: download last night's copy, restore it into a temporary database, count the tables and rows. Identical to production, for all three databases. I'll redo that test every quarter.

For the record, the three backups weigh under 200 KB in total. The storage's free plan takes 10 GB.

Knowing before your customers

The worst case is a customer writing "your site is down" when you had no idea. Three free tools prevent it:

  • outside checks call each of my sites every 5 minutes. At the slightest problem I get an e-mail, and another when the site comes back;
  • the server itself warns me about a failed deploy, a nearly full disk or an unreachable machine;
  • error tracking gathers errors (on the server and in visitors' browsers) and the last 30 days of logs, with an e-mail for each new important error. I configured it to keep nothing personal: no identity, no IP address, no form content.
Uptime checksevery 5 minutesThe appson the VPSFallback pageif the server is downError trackingerrors + logs, 30 daysMy inboxalerts, before users notice
Outside checks call the sites every five minutes, errors and logs land in one place, and anything wrong reaches my inbox — not a customer first.

Why outside checks? Because a tool installed on the server goes down with it. The fire alarm shouldn't be in the room that's burning.

Plan B: one server, so what?

One server is one point of failure. But there are several levels of answer, from free to paid:

  1. The data lives elsewhere. The hardest part to duplicate, the database, already isn't on the server. If it disappears, you restart the apps elsewhere.
  2. A written rebuild plan. New server, security, deploy tool, apps, nightly tasks: about an hour, following a checklist rather than memory.
  3. A fallback page. If the server stops answering, Cloudflare shows visitors a clean "back in a few minutes" page instead of a raw error. Payments and logins, on the other hand, get the real error, so they know to retry.
  4. A real second server, with automatic failover. That's about ten more euros a month. I'll do it the day an app earns enough to pay for it.
→The right level, at the right time

Perfect redundancy for a project with no customers yet is money spent too early. What matters is knowing exactly what happens if the machine goes down, and how long it takes to come back.

People's data

My apps store e-mail addresses, names, CVs and names on certificates. That's personal data, and two laws apply to me: GDPR, because my users can be in Europe, and Cameroon's data protection law, in force since June 2026, which applies to my business.

In practice:

  • honest privacy policies, naming every provider, saying where data lives and how long it's kept;
  • cookie-free analytics that only see totals;
  • a record of processing: for each use of data, who is affected, why, for how long, and where;
  • a written data-breach procedure: contain, understand, assess, notify the authority within 72 hours if needed, tell the people affected if the risk is high, and log everything.

What the AI did, what I did

I didn't do all of this alone: an AI assistant wrote a good share of the scripts and configuration. But the viral post was right about one thing. The developer answers for what they ship. So at every step:

  • I asked why before asking how;
  • I checked myself whatever could be checked: tests, logins, restored backups;
  • and I wrote it all down, up to this series, so I can explain it without notes.

If you keep only one thing from these five posts: a server isn't defined by the moment it starts. It's judged on the day something breaks. Mine is ready for that day. And if you have questions, write to me.