Skip to content

Maintenance, Support & Optimisation

Neon free tier compute hours and a health check

A five-minute uptime check kept Easinvy Retail's Neon database awake: 19.4 CU-hours in 78 hours, and the free allowance gone on 30 September 2026.

Ridam Agrawal

Founder, Easinvy

Published
Read
6 min
On this page

In short

Does a health check keep a Neon compute awake?

Yes, if it queries the database at least every five minutes. Neon's Free plan suspends a compute after five idle minutes, and Free users cannot change that. A five-minute uptime check on Easinvy Retail kept its 0.25 CU compute awake around the clock: about 6 CU-hours a day, which spends the free 100 in under 17 days.

Easinvy Retail, our billing and inventory product for shops, keeps its data in PostgreSQL on Neon's Free plan. From 14 September to 14 October 2026 an uptime monitor checked the backend every five minutes, and the check ran a query against the database. Neon suspends an idle Free compute after five minutes, so the compute never went idle, and the month's free compute ran out on 30 September. Every figure below comes from Neon's own alert emails, the monitor's emails and the backend's git history.

The health check that asked the database

Easinvy Retail's health route was built to catch the outage that matters. On 14 September 2026 the backend's /health route was changed to run select 1 against the database, with a five-second limit, and to answer 503 {"ok":false,"db":"down"} when that failed. A process that is up while its database is unreachable is exactly what a monitor exists to catch, and a route that answered from memory would have hidden it. The five seconds were chosen to give Neon time to wake a suspended compute after a quiet night.

The same day, three UptimeRobot monitors went live. The one on /health was a keyword check on "db":"ok", tested against a real failure: with the backend pointed at an unreachable database, the route answered 503 and the check went red. The /health monitor ran every five minutes.

Five minutes against five minutes

Neon's Free plan suspends a compute after five idle minutes. Neon's scale-to-zero documentation says a compute "scales to zero after an inactive period of 5 minutes", and that for Free plan users "this setting is fixed." A query every five minutes lands before the compute can go idle, so Easinvy Retail's compute had no quiet night to wake from. The health check that was meant to tolerate a cold start prevented every one.

A controlled test by DevOps Daily found the same edge: a SELECT 1 every five minutes kept a test compute awake for 100% of a 40-minute run, every six minutes 90.6%, and every ten minutes 52.1%. Its monthly figures are projections, which it says plainly. Easinvy Retail's figures below are Neon's own meter on a live project.

What Neon's meter recorded

Neon's alert emails read the meter twice in September 2026. At 20:16 IST on 26 September, Easinvy Retail's project had used 80.6 CU-hours of its 100 CU-hour monthly allowance. At 02:10 IST on 30 September it had used all 100. That is 19.4 CU-hours in 77.9 hours, an average of 0.249 compute units in every one of those hours. Neon counts compute as size in CU times hours running, and 0.25 CU is the Free plan's smallest size, so the meter matches a compute at its smallest size running without a break for three days.

At that rate a compute uses 6 CU-hours a day. Neon's plans page puts the free allowance as "enough to run a 0.25 CU compute in a project for 400 hours/month". Four hundred hours is 16.7 days, so a compute that never suspends uses the free allowance in under 17 days of any month. October 2026 repeated September: Neon recorded 80.1 CU-hours by 10:46 IST on 14 October.

The day Neon's free allowance ran out

Neon's email at 02:10 IST on 30 September 2026 said that "Apps for this project may fail to connect" and that queries "will error until your allowance resets next month". Neon's plans page says the same in its own words: the compute is suspended until the next billing period or an upgrade, connections fail with SQLSTATE 53000, and "None of these limits delete your data."

Easinvy Retail kept answering for another eleven hours. At 13:06:26 IST the monitor's keyword check failed, and at 13:11:30 it passed again: 5 minutes and 4 seconds. Nothing was committed to any of our repositories between 29 September and 1 October, so no release explains it, and we did not establish why the outage lasted five minutes rather than until the allowance reset. The monitor's September report scored the database check at 99.988%. The monitor that spent the allowance was also the only thing that saw the outage it caused.

The warnings Neon sent on the Free plan

Neon emailed three times about Easinvy Retail's compute: at 81% (80.6 CU-hours) on 26 September, at 100% on 30 September, and at 80% (80.1 CU-hours) on 14 October 2026. Each came from Neon's alerts address and named the project and the allowance. Neon's plans page lists spending notifications as not available on the Free plan; those are alerts about charges. The compute warnings arrived on Free all the same, and on this project they were the only warning that the database was about to stop. The inbox they go to has to be one somebody reads.

The fix, thirty minutes after the 80% email

Neon's 80% email arrived at 10:46 IST on 14 October 2026, and the fix was committed at 11:16 IST. The backend gained a second health route, /health/live, which answers from the process alone and never touches the database, so a monitor can poll it as often as it likes. The frequent check was to move there, and /health, the database check, to run hourly at most.

Twenty-five minutes later we went further and deleted the monitors. At Easinvy Retail's size, a shop reporting a problem is soon enough. That choice has a cost, written into our own runbook: nothing pages anyone when the database is down. What remains is a rule written above the route, "never poll it faster than hourly", and Neon's own emails for the allowance.

After the fix

Neon's console shows what the change bought. On 14 November 2026, Easinvy Retail's project had used 26.39 CU-hours since its billing period began on 1 November: about 2 CU-hours a day over 13 days, against 6 a day while the monitor ran. At that pace November uses about 60 of its 100 free CU-hours. The project's monitoring chart now shows the compute suspending between uses, which September's meter shows it never did. The project still lets the compute scale from 0.25 to 2 CU, so the 26.39 CU-hours measure what the compute cost, not how many hours it was awake.

A liveness check that never touches the database

Easinvy Retail's backend now has two health routes, and each has one job.

RouteTouches the databaseUse
/health/liveNoThe process is up and routing. Safe at any interval
/healthYes, select 1The database answers. Hourly at most on Neon's Free plan

A liveness route answers whether the server is running, and costs nothing on a database that scales to zero. A readiness route that queries the database answers the question that matters more, and on Neon's Free plan every call to it can wake the compute for five minutes. Spacing the database check an hour or more apart lets the compute suspend between checks. Neon also reports a compute's state through its API, and its network transfer page says that "Consumption API calls do not wake suspended computes, so polling does not increase your compute usage." Easinvy Retail connects to Neon through its connection pooler, as described in Multi-tenancy in Drizzle, from a shipped product.

A checklist from this incident

  1. Know your database's idle timeout before choosing a monitor's interval. On Neon's Free plan it is five minutes and fixed.
  2. Point the frequent check at a route that never queries the database.
  3. Run the database check an hour or more apart, or not on a schedule at all.
  4. Read the allowance as hours: 100 CU-hours is 400 hours at 0.25 CU, less than a month of an always-on compute.
  5. Keep Neon's alert emails going to an inbox somebody reads. On the Free plan they were the only warning.

Easinvy's maintenance and support work includes operations problems like this one.

Questions people ask

How long do Neon's free compute hours last?
Neon's Free plan gives each project 100 CU-hours a month. A compute at the smallest size, 0.25 CU, that never suspends uses 6 CU-hours a day, so the allowance lasts about 16.7 days, or 400 hours. A compute that suspends whenever it is idle lasts far longer.
What happens when a Neon free project uses all its compute hours?
Neon suspends the compute until the next billing period or an upgrade, and connections fail with SQLSTATE 53000, though no data is deleted. On Easinvy Retail, Neon's email came at 02:10 IST on 30 September 2026, and the health check failed for five minutes that afternoon.
Does an uptime monitor keep a Neon database awake?
Only when its check reaches the database. A check that runs a query every five minutes or more often keeps a Free compute awake, because Neon suspends after five idle minutes. A check that only asks whether the server process is running never wakes the database.
Is Neon free to use?
Neon has a Free plan with 100 CU-hours of compute per project per month, enough for a 0.25 CU compute to run for 400 hours. When a project uses all of it, Neon suspends the compute until the next billing period or an upgrade, and the data is kept.

Sources

Ridam Agrawal

Founder and engineer at Easinvy. Writes up the parts of building and running a multi-tenant product that were expensive to get wrong. More about the studio →

Keep reading