Archives

After pointing a domain at a new server, one machine kept loading the old site even after flushing DNS. The stale record was in the ISP's resolver, not on the machine. Here is how to prove that in three commands.

After switching nameservers to Cloudflare, 16 of my self-hosted subdomains stopped resolving. The apps were fine. The A records had not been carried over.

How I moved my apex domain from Vercel to a self-hosted Next.js container: rebuild as indexable, add an nginx vhost, change one A record, run certbot, keep a noindex mirror.

My VPS had about 2.4 GB of free RAM and no swap. I added a 4 GB swapfile for the Next.js build, removed it afterwards, and later moved the build off the server.

The workflow I use to build an image on GitHub, push it to GHCR and restart the container on my VPS over SSH, with the two things that made the first runs fail.

An open-source Gradio app defaulted to CUDA and NVENC. I packaged it for a VPS with no GPU: CPU torch wheels, a baked-in model, a login and a memory cap.

How I built a private notes app: FastAPI, SQLite for text, a self-hosted S3 bucket for files, and a storage key that can only touch that one bucket.

A root redirect in nginx broke sign-in to the RustFS console, and a blanket rate limit returned 429s right after login. Two nginx map blocks fixed both.

I needed S3-compatible storage next to n8n on one VPS and did not want to install MinIO. RustFS in one container, behind nginx, uses about 60–90 MB of RAM.

A four-node n8n error workflow that sends failure emails through the Gmail API, plus the patterns I use so failures clean up after themselves and carry a useful message.

I posted to an n8n hosted form from a script and every value arrived as null. The form expects multipart keys field-0, field-1, not the labels. The reply is also not the final result.

n8n 2.x excludes the Execute Command node by default, so my workflow would not activate. I left it disabled and moved the shell work into a small HTTP sidecar on a private Docker network.

I build seven n8n workflows from one Python file and push them with the n8n public API. Here are the helpers, the deploy calls, and the expression bug that cost me the most time.

How I run headed Playwright tests on a headless server and watch the real browser in a web page: Xvfb, x11vnc and noVNC in one container, locked behind nginx auth_request.

The Docker flags, mounts and resource caps I use to run an AI agent app with shell access on a shared 2 vCPU VPS, plus the localhost-bind problem that needed a one-line patch.

A working GitHub Actions workflow that boots a hardware-accelerated Android emulator and runs instrumented tests in about three minutes, and why AVD snapshot caching made it worse.

I put the open-source mcp-gsc server on a VPS in a locked-down container, gated it with an nginx bearer token, and hit a map hash error that broke nginx for every site.