Hey dev.to — Lepp again. Entry two in building Neuro-Kindred (a free, calm social app for neurodivergent adults, 18+) in the open. Last time: pebbles instead of likes. This time: a mystery, three wrong suspects, and the week I nearly lost the whole app to a free-tier limit.
The setup
Neuro-Kindred runs on a $3/month server and a free Postgres database (Neon, 100 free compute-hours a month). On Nov 15 I got the email nobody wants: 80 of the 100 hours already used, halfway through the month. At 100, the database sleeps until December 1st — app down, no charge, data safe, but down.
Something was waking the database constantly. It wasn't my members: in the worst 48-hour stretch, 38 members sent 63 messages total. So I went hunting.
Suspect #1: the scanners
The server logs showed a burst of 188 requests in one minute from a cloud IP, probing for .env files and admin paths. All 404s, nothing exposed — but bots were my first theory. Verdict: real, but too small. The burn continued on quiet nights with no scanner traffic at all.
Suspect #2: the old deployment
The app used to live on an app-builder platform before I moved it to my own server. Maybe the old copy was still running somewhere, pinging away? I unpublished it (Nov 17) and measured for 24 hours. The burn barely moved. Not the culprit — and now I couldn't easily un-unpublish my main suspect.
Suspect #3: my own forgotten robots
This one stung. Back in September, the builder era set up two keep-alive jobs (duplicates, of course) pinging the app's health endpoint every 5 minutes — and every ping runs database queries, and every query resets the database's 5-minute idle timer. The database was being politely woken up 24/7 by robots I owned. I disabled both (Nov 15) and the sleep test passed... but the daily burn only dropped a little. A contributor. Not the culprit.
The actual culprit: the app itself
The final measurement read the request logs by endpoint. In 48 hours: the calls screen had been checked 3,497 times — about once every 3 seconds — from one or two phones with the page left open. Message threads every 4–5 seconds. Game rooms every 2 seconds. My own app's screens were the waker. One member leaving the app open on their phone was costing more database than every real conversation combined.
The fix shipped Nov 20: the calls overview now checks every 20 seconds, message threads every 10 (pausing entirely when the tab is hidden was already in there — the problem was never hidden tabs, it was visible ones being impatient). Full test suite green (552 tests), and I'm re-measuring for 24 hours before I claim it worked. If the number doesn't drop, I'll say so in entry three.
The honest numbers
- Database: 86.3 of 100 free hours used (Nov 18 reading)
- Members: 38 · Revenue: $0 · Server: $3/month
- Cost of the investigation: a few evenings and one unpublished old site
Lesson: on a free tier, your own polling is your biggest user. Measure by endpoint before you blame the internet.
One question before you go
Neuro-Kindred runs a community night every Friday, 7pm CT (trivia, a game, then a watch-together on Discord) — it's the hour the app feels most like itself. Comment with the interest you could talk about for a full hour, and I'll send you an invite to this Friday's. That's the whole exchange — your weird topic for a seat at the table. It's also literally what the app is for.
— Lepp
Next entry: "Pebbles grow old now" — why an unanswered pebble sitting in limbo forever is unkind, and the 30-day close we just shipped.