The six
| App | What Supabase does in it |
|---|---|
| Excel test-prep site | Password and Google sign-in, saved progress |
| Dating platform with ID verification | Logins, 18 tables, photo and ID-document storage |
| Screenshot API | Magic-link login, hashed API keys |
| Website technology detector API | Login, API keys, a 7-day results cache |
| Tweet-writing mobile app | Logins, database, file storage |
| Waitlist and founder-seat sales for a LinkedIn tool | One table with an atomic seat counter |
None of these are AI products at the database layer. Supabase is on this list because the systems I build still need accounts and permissions, and this is the piece I stopped rebuilding.
Trap one: free projects pause
A free Supabase project that gets no activity for a week is paused. A paused project looks like a total outage. Logins fail and the app shows nothing, and someone has to click a button in the dashboard to bring it back.
Pre-launch apps go quiet for a week all the time. So every project I create gets a keepalive the same day: a systemd timer on my server that runs one trivial query every three days. The setup is a template, one small env file per project. On November 13, 2026 I added the timer for the Excel site, a week after starting it, which was later than my own rule allows.
The free plan also limits you to 2 active projects. Past that you are paying or pausing.
Trap two: the login that hangs until you refresh
On the dating platform, users would sign in and get a spinner that never ended. Refreshing the page fixed it. The cause was inside the Supabase JavaScript client, version 2. It takes a browser Web Lock around auth operations, and under some sequences the lock never released.
Two changes ended it:
- Pass a no-op lock when creating the client:
auth: { lock: (_name, _timeout, fn) => fn() }. That turns the lock off. The risk is two tabs refreshing a session at the same instant, which I accept. - Never call
getSession()when the app mounts, and never run a second database query inside the post-login handler. LetonAuthStateChangedeliver the first session and fetch the profile once.
I now diff the built login code against the previous release before every deploy of that app. If those files changed and I did not mean to change them, the deploy stops.
Trap three: the key that was not a key
On November 11, 2026 I found the Excel site's API returning 401 from Supabase. The service-role key in the server's env file was wrong. I had copied it from the output of the Supabase CLI command that lists a project's keys, and that output is masked. It looked like a key. It was a redacted placeholder.
Copy service keys from the dashboard, and have the server make one authenticated call at startup so a bad key fails loudly on deploy day instead of quietly a week later.
Row-level security: on or off, nothing between
Supabase exposes your database straight to the browser, guarded by Postgres row-level security policies. With RLS off, anyone holding the public key can read the table.
On the dating platform, RLS is on for all 18 tables, public and logged-in roles are denied by default, and only the server's service role can write sensitive rows. Photos live in a public bucket capped at 5 MB a file. ID documents live in a private bucket.
The Excel site has a test file that runs against the policies themselves.
Free versus Pro, as listed November 14, 2026
| Free | Pro, from $25 a month | |
|---|---|---|
| Database | 500 MB | 8 GB, then $0.125 per GB |
| Monthly active users | 50,000 | 100,000, then $0.00325 each |
| File storage | 1 GB | 100 GB |
| Egress | 5 GB | 250 GB |
| Backups | None | 7 days |
| Pausing | After 1 week idle | Never |
My rule: free tier with a keepalive until the first paying customer, Pro the same day. The reason is the backups row. A free project has none, and a keepalive does not protect you from a bad migration.