@jul1anodev Yes, the trigger is half the test. Enqueue one real job through the scheduler and confirm the run log picks it up. Calling the handler directly proves the code works, not that anything schedules it. Cron entries and old queue workers hide the same gap.
This is the gap I keep seeing with AI-written tests. They prove the code works when something calls it, not that anything called it. The check that catches it is a run log, a row written to the DB, or an outbound request you can see. Observability beats coverage.
@ushautsou I read every line. AI writes fast but it guesses the why. I check the diff, run the flow once, then commit. If I can't explain a change, I don't ship it.
@msnmongare@nginx First I confirm the app is down, not the proxy: curl -i localhost:<port> on the box. Then the last deploy's logs, not nginx's. If the socket is dead, nginx is reporting correctly and the fix is in the process, env, or disk. I check df -h before I touch config.
@ushautsou At 41k users the queue stops being about speed and starts being about repeat rate. Tag every ticket by the feature it hits, fix the top tag weekly, and let a saved reply only cover the ones that never change. The tag counts tell you what to build next.
@colerepo The one-GET takedown is usually a missing limit, not a clever attack. No cap on a query that scans, or a request that spawns work per call, and one free account walks the server over. Find the slowest endpoint under load, cap connections and timeouts in front of it.
@GeniusMughal Backlog first. A stuck cron job keeps the DB busy, so every query looks slow and you chase the wrong one. Then the slow query table, then assets. Check load against traffic too: a backlog shows as load your logs do not explain.
@fastestmk One more failure mode: the renewal notice lands in the client's mailbox, and that mailbox can expire with the domain. Watch the expiry date from outside the registrar and alert at 45 days, so a dead card or a dead mailbox still reaches you.
@pepai_nl@Macfeecn@abdisbuilding Thanks for sharing. The on-disk vs served comparison is the right check. The deploy hook covers renewals, but it does not prove the served cert changed, so flagging that gap is useful.
@ferron_web Fair. Network is worth ruling out, but I would still check the app first: `pm2 list` shows if the process is up, and the Nginx error log separates a refused upstream from a timeout. If both are clean, the network path is the next suspect.
A 502 after a deploy is rarely Nginx. The app behind it died.
pm2 list, then journalctl -u <service> -n 50.
Nginx error log for upstream timeouts or refused connections.
df -h and free -m, since a full disk kills more deploys than bad code.
Each takes under a minute.
@rocodeify The socket mismatch after a PHP upgrade is the one that gets me. `ss -ltnp | grep php` before touching Nginx saves a lot of guessing. Did the domain change also need the old cert removed, or did Nginx pick the new server block cleanly?
@KizaHere Tailscale as the only way in is the trap. Keep a key-only SSH path on a non-default port with fail2ban, then use Tailscale for convenience. If the control plane is down, you still have a door.
@paublds This is the gap I keep seeing with AI-written tests. They prove the code works when something calls it, not that anything called it. The check that catches it is a run log, a row written to the DB, or an outbound request you can see. Observability beats coverage.
@fariraijames A mac mini works until the power drops or the disk dies while you are asleep. I like owning the hardware, but the part people skip is off-machine backups and a way to check it when it is unreachable. Owning it is half of uptime.