Nothing teaches you about a system like watching it fail. A few favourites:
The time I took everything down myself. To keep one slow page from holding up everyone else, I moved page rendering onto a pool of worker threads. Then, to check that site moves had no hiccups, I sent a flood of test traffic at the live sites. They froze for nine minutes and stayed frozen after the traffic stopped. Pages were each holding one database connection while asking for a second, and with dozens of threads and a pool of fifteen connections, every thread ended up waiting for another one. Meanwhile the health check, which never touched the database, cheerfully reported that everything was fine.
The fix: database pools that can't deadlock, a limiter that turns an overload into a quick "try again" instead of a hang (with a fair share for each site, so one busy site can't starve the others), a health check that notices a stuck process, a watchdog that restarts it and a firm rule about where load tests happen.
The 45-second hiccup. Most site moves were seamless, but one kind wasn't: if the move changed the site's routing rules, the old and new copies briefly described the same route in two different ways. The front door, unable to decide which to believe, served neither for about 45 seconds. Every other part of the move was correct, which is exactly why it took a while to spot.
A tiny site at 100% CPU. A small shop with a few hundred visitors a day kept maxing out its processor. The culprit wasn't customers, it was automated scanners firing more than a hundred requests in ten seconds, looking for files like /.env or /wp-login.php that a FolioPress site never has. Each one got a beautiful, fully designed "page not found": about 70 ms of work, a hundred times over.
The fix: those requests now get a nine-byte "Not found" before any real work happens, and Cloudflare turns most of them away before they arrive.
"Server not found" for a site that was up. A brand-new site's web address is created during setup. If anything looked the name up a moment too early, internet providers remembered "this name doesn't exist" for up to thirty minutes, even though the site was working fine. The fix was a catch-all address, so every name under foliopress.io always exists; names without a site get a friendly "there's no site here" page.
Neighbours. Hosting many sites side by side raises questions a single website never has to answer: what one site's code can do to another, and what a site can claim that isn't its own. A security review before opening sign-ups turned up a few of these. All of them are fixed: stricter cookies, proof of ownership for custom domains, and every site mounted in its own separate folder, so no site can reach another's files.