My Blog Returned 502. The Container Engine Was Stopped.
I opened a publishing session on November 15 and couldn’t list my WordPress posts. The command failed before it reached WordPress: Docker couldn’t connect to its configured socket.
The public homepage also returned HTTP 502. My local WordPress port refused connections. This wasn’t just an administrative command failing while the public site continued working normally.
I recovered the service by starting the existing OrbStack engine. I still don’t know why it had stopped, when the interruption began, or how many visitors were affected. A successful recovery didn’t fill those gaps.
The dependency I had stopped noticing
This blog runs WordPress and MariaDB in Docker containers on a Mac. OrbStack provides their container runtime. Cloudflare connects public requests to that local service through a tunnel.
The public domain doesn’t make the local runtime optional. If WordPress isn’t listening, having a domain and an edge connection doesn’t manufacture a working origin server.
OrbStack’s documentation describes its Docker engine, networking, and persistent volumes. In my setup, WordPress serves pages and MariaDB stores the site data. Both depend on the engine being available.
I hadn’t lost the distinction on paper. I’d lost it in my routine: I approached the machine expecting a publishing task, and found an unavailable service instead.
Four observations, one recovery
I didn’t diagnose the stopped engine from the 502 response alone. I checked separate parts of the path.
| Check during the incident | Observed result | What it established |
|---|---|---|
| WordPress command through Docker | Configured Docker socket unavailable | That command couldn’t reach the engine |
| Local WordPress connection | Connection refused | The expected local service wasn’t listening |
| Public homepage | HTTP 502 | That public request failed |
| OrbStack status | Stopped | The engine reported that it wasn’t running |
| After engine startup | Containers running; public HTTP 200 | Service recovered at the tested endpoint |
The OrbStack application process was present during diagnosis. That wasn’t proof that its engine was running. Opening an application and checking its runtime state answered different questions.
I started the existing engine rather than creating a replacement database or rebuilding the site. The existing WordPress and database containers came up. I could list the posts again and confirm the public homepage response.
I didn’t change the tunnel configuration, container definitions, or stored content to achieve that recovery. I also didn’t establish that a particular update, user action, or resource limit caused the stop.
Why I won’t call every 502 this problem
Cloudflare’s troubleshooting documentation covers several sources of 502 and 504 errors. The response code doesn’t uniquely identify a stopped container engine.
In my incident, the local refusal and stopped-runtime status supplied the additional evidence. On another machine, I’d need to check that machine’s path rather than repeat my recovery command as a universal fix.
I haven’t measured an exact outage duration. The first failed request I observed marks when I discovered the problem. It doesn’t mark when the problem began.
I also haven’t proven that every page failed for every visitor. A browser or edge cache could have served an already stored response. I didn’t inspect per-visitor cache behavior during the interruption, so that remains a possibility, not a count of successful visits.
Recovery isn’t prevention
The immediate cost was concrete: the homepage request failed, and publishing couldn’t proceed until the runtime was available. I can’t honestly turn that into a lost-reader or lost-revenue estimate.
I’d want an external availability check to notice this independently of my next publishing session. A local runtime check would then help identify which dependency failed. Those are different checks with different jobs.
I haven’t installed a new watchdog or alerting system as part of this article. The fact that I now understand the dependency doesn’t mean unattended recovery is implemented. I’d need to verify that behavior separately.
Don’t read the silence as declining interest
An unavailable origin can stop producing normal page-view records even while people are trying to visit. Cached deliveries can also bypass that origin. A quiet log therefore isn’t enough to conclude that readers stopped searching for my articles.
That matters because I’d recently used article-path log analysis to guide my next topics. Availability belongs beside those traffic windows. I can’t use an unknown outage interval as a clean editorial comparison.
My November 19 homepage check returned 200 again. That’s a current observation, not proof of uninterrupted service since recovery. For now, I have a working blog, a documented dependency failure, and a cause I haven’t established.