Ninety days ago I pointed a real SaaS — a booking tool with ~1,800 users and ~60k queries/day — at Neon's free tier and deleted my $25/month VPS Postgres. My card has been charged $0.00 since. But two free-tier limits came within a hair of taking the site down, and nobody's landing page mentions either.
The workload, by the numbers
| Metric | Value |
|---|---|
| Database size (day 90) | 412 MB (limit: 512 MB — I'm at 80%) |
| Queries/day | ~60k |
| Compute hours used/month | ~178 (limit: 191.9 — I'm at 93%) |
| p50 query latency | 9 ms |
| Cold start after idle | 340-480 ms |
| Downtime incidents | 1 (self-inflicted, see below) |
Limit #1: compute hours are a monthly budget, not a throttle
Neon's free tier gives you 191.9 compute-hours/month on a 0.25 CU baseline. An always-on branch at minimum size burns ~180 hours/month just existing. My first month, a runaway analytics query loop kept the branch active through idle periods and I hit 187 hours on day 26 — the branch auto-suspended and stopped accepting connections. The fix was a cron job that forces suspend at 85% budget, which is a sentence I can't believe I'm typing about a "serverless" database.
Limit #2: 512 MB is smaller than you think
Postgres bloat is real. My events table's indexes alone are 96 MB. VACUUM discipline went from "nice to have" to "weekly cron" — pg_stat_user_tables showed 31% dead tuples before I got serious about it.
Cold starts are fine — mostly
The 340-480ms wake-up is invisible for a booking flow (users tolerate one slow first click). It would be disqualifying for a real-time API. I sketched the latency harness with MonkeyCode's free tier: https://ly.cyberserval.tech/iIETXiF
The controversial part
Everyone says "free tiers are for prototypes." After 90 days I disagree — with caveats. Neon free tier is production-capable if your app is bursty, under 400 MB, and you automate the compute-hour budget. It's a trap if you're always-on with a chatty ORM, because the budget math collapses the moment your branch never sleeps.
Which free-tier limit has actually bitten you in production — storage, compute hours, or cold starts?
Top comments (0)