A travel search on a few of my sites took about six seconds every time someone started fresh, and I’d have sworn it was cached. It was, on the way in. The writer stored the key without a trailing underscore; the reader asked for it with one. A forked sibling plugin had the opposite bug, a doubled underscore, plus a read that had been commented out altogether. The fix is one function that builds the key for both sides, then a daily cron that warms the cache inside a time budget. A cache entry existing only proves the writer ran. The only proof a cache works is a hit.

Pick a currency we had no conversion rate for and the page swapped the symbol but left the number where it was. A cabin at 1,200 in the base currency displayed as £1,200. No error, no log line, no visual tell, and a completely plausible price for a cruise cabin. Two functions, each correct on its own terms: the label side honoured the visitor’s cookie unconditionally, the maths side sensibly refused to convert without a rate. The fix is the weakest guarantee worth having — if we can’t convert, we don’t relabel. The shape to hunt for in your own code is a shared input whose validity only one side ever checks.

Four model calls running side by side instead of one long one, and the wall clock went up. Not the calls — those were faster, exactly as advertised. The time was in the gaps between stages. I’d ranked Action Scheduler as the preferred dispatcher on reliability grounds, and on a quiet site it runs its queue off WP-cron, which runs when somebody loads a page. Three handoffs at about a minute each, bought to save ninety seconds of generation. The fix was a reordering rather than a rewrite, and only because an atomic claim I’d written months earlier for a failure case turned out to be load-bearing as the design.

php -m listed imagick. Site Health, on the same box, insisted it wasn’t installed. One of them was lying. PHP on a web server isn’t one thing — CLI, Apache and FPM are separate SAPIs with separate conf.d directories, and phpenmod without -s ALL only wires up the SAPIs that exist at the moment you run it. This box was provisioned in the mod_php era and gained FPM later, so the one SAPI actually serving pages never got the symlink. Nothing errored; it just wasn’t there. Fix is -s ALL plus a version-discovered FPM reload. When something swears an extension is missing, work out which PHP is doing the complaining.

A cruise package that departed on a Saturday was showing Day 1 as the Friday. Not every package, which is what made me distrust my own eyes for the first hour. The cause was a date constructed with no timezone: PHP quietly uses the server’s, the rest of the feed handling assumed UTC, and a few hours of offset is all it takes to land on the wrong side of midnight. Pin the timezone where you build the date and make every date in the flow agree. I had already “fixed” this once in July and called it minor in the commit message. It was not minor. It was quiet.

A few weeks after writing that CURLOPT_TIMEOUT of zero is never the right value, I typed it in myself. An itinerary build streams for several minutes with live searches in the middle, and a 300-second total cap was cutting off a stream that was still delivering. A total timeout can’t tell working from dead — both look like still open, still not finished. So measure silence instead: no total cap, and a low-speed watchdog that gives up after two minutes below a byte a second. That only works because the upstream sends pings while it thinks, so check before you lean on it. Match the timeout to the thing you’re actually measuring.

A migrated multisite where three of the five subsites sit on their own external domains. Front end perfect everywhere, login dead on the mapped three, and WordPress blaming the browser. It was never the browser. With no COOKIE_DOMAIN defined, WordPress stamps every auth cookie with the network’s primary domain, and no browser will accept a cookie scoped to an unrelated apex. Authentication succeeded; the cookie simply wasn’t allowed to land. One constant, scoped to the host the request arrived on, fixes it — at the cost of network-wide single sign-on. The old server had this hand-patched years ago. A migration copies files and databases, not undocumented config edits.

Two JSON files cached side by side from the same feed, pulled by the same importer, using id and parentId in opposite senses. Nothing in either file tells you which one you’re holding. That single wrong assumption produced two failure modes: an admin dropdown that threw a TypeError the day it appeared, and a second one that had been silently collapsing twelve hundred ships into one row per cruise line for years. The loud one got reported in an afternoon. The quiet one just looked like a short list, and nobody counts the options in a dropdown. There’s a third copy of the same loop still cached on the front end.

The site wasn’t down, which was the confusing part. It answered, forty seconds at a time, in waves. The PHP-FPM slow log pointed every stuck request at the same curl_exec(), talking to a travel API that was having a bad afternoon. Our cURL options carried CURLOPT_TIMEOUT set to zero, which doesn’t mean no time — it means no limit. Slow calls held FPM workers until the pool filled, and then pages that never touched that API queued up behind them too. The fix is a total timeout, a connect timeout, and actually handling the error when it fires. Leave it at zero and you’ve handed a stranger the power to freeze your site by being slow.

On EC2 you’d attach an IAM role and never see a key. Lightsail can’t do that, and during the migration both servers needed to reach S3. The setup: a dedicated IAM user with a tightly scoped policy, keys kept on the operator laptop only, and credentials handed over as environment variables for a single SSH command, never written to disk. Writing it up, I found a few things I’d got wrong, not least that the servers were getting far more permission than they needed. The fix is fifteen-minute, single-bucket credentials from STS.