4 min read

Product

My Deploy Tool Returned HTTP 200 for Three Weeks and Published Nothing

Scheduled posts on this blog sat unpublished while every log line said success. The deploy API answered 200 each night and quietly skipped the build.

coolify deploy http 200 no buildcoolify force=falseastro scheduled posts static build

My Deploy Tool Returned HTTP 200 for Three Weeks and Published Nothing · Products Decoded
By Debashis Nayak · Products Decoded10 Nov 2026

I had a cron job that rebuilt this blog every night, and its log was a perfect wall of deploy HTTP 200, three weeks long. In those three weeks it never built anything.

I only found out because scheduled posts kept showing up late. A post dated for one day would appear on whichever day I next pushed code, which is a very different thing from appearing on its own date.

Why a static blog needs a nightly rebuild at all

First, the part that works as designed. This site is built as static output. The blog’s page routes are generated at build time, and the code that decides which posts exist filters out anything dated in the future. That filter runs when the build runs, not when a visitor arrives.

So a post dated for tomorrow is simply left out of the build. It returns a 404 until some build happens on or after its date. Nothing in the site changes at midnight by itself, which is why I set up a daily cron to trigger a rebuild.

The cron existed from 27 August. It fired at 00:30 UTC every day and called my deploy tool’s API.

The bug: force=false

The tool is Coolify, which I use to deploy my sites. My script called its deploy endpoint with force=false.

With that flag, Coolify compares against the last commit it deployed. If the commit has not changed, it skips the build. It then returns HTTP 200 anyway.

That is a reasonable thing for a deploy API to do if you read the 200 as “request accepted”. I read it as “built and shipped”, because that is what a green log line trains you to believe. The log in my case showed an unbroken run of deploy HTTP 200 and no build had run behind any of them.

Posts still went live eventually, because I push code often and any push triggers a real build. That is exactly what hid it. The blog looked like it worked, just a bit unpredictably, and I put the lateness down to my own pushing habits.

How I proved the cron never built anything

Logs could not settle this, since the logs were the thing lying. I needed something the cron could not have faked.

Coolify tags build images by commit SHA. If nightly builds were running, I would expect fresh images at 00:30 UTC, even with an unchanged SHA. On 17 September there were only three images for this app, created at:

Image createdMatches
09-12 07:12a git push
09-16 07:23a git push
09-17 06:44a git push

Every one landed at a push time. None at 00:30. The container creation times matched the same pushes. That was the proof: three weeks of nightly 200s, zero nightly builds.

The fix

I changed force=false to force=true in the script that installs the cron, then reran the installer. It is idempotent, so it replaced both the remote script and its cron line. I kept a backup of the previous version first, and I checked disk before agreeing to a forced daily build. There was plenty free, 434G, so one build a night costs me nothing that matters.

At the time I wrote this up, one thing was still open. The reinstall returned HTTP 200 again, which by now meant nothing to me, and I could not confirm whether a fresh image appeared for an unchanged SHA. A permission block stopped that check. So I set myself a test with an observable outcome instead. A post called drafts-yes-sends-no was dated 18 September, and it had to be live on this site after 06:00 IST that morning. If it was not, force=true had not taken. My fallback plan was to make the blog pages render on request instead of at build time, which removes the need for any rebuild.

After being burned by a 200, I was not going to accept the next one on trust.

One curl is not proof of a deploy

The same day I hit a second version of the same problem. Right after a push, the new cover image for a post answered 200 on one request and 404 on the very next. My read is that the old and new containers were briefly both answering behind the same route, with requests alternating between them.

So a single successful check is a coin flip during that window. Two ways around it:

  • Sample the URL around ten times and look at all the answers.
  • Gate on the number of containers carrying the proxy’s labels for that app. When it is exactly 1, the old one is gone and what you see is the new build.

Questions I now ask of any “green” automation

  • What artifact does a real run leave behind that a skipped run would not? An image with a new timestamp, a file with a new hash, a post that is actually live.
  • Does the success code mean “accepted” or “done”?
  • Have I ever watched it fail? A check that cannot fail cannot reassure me.
  • Am I sampling once or enough times to rule out two versions answering at once?

I wrote about another quiet failure with a clean dashboard in the server bill that kept coming after I deleted the server. Different tool, same shape. The status said everything was fine, and the evidence sat one layer down, in a place I had not thought to look until the output stopped matching what I could see.