Watchtower

Database and cache. Every poll runs a real query and a cache write and read, and reports how long each took.

Queue workers. A heartbeat job goes through the queue every minute, so Watchtower knows a worker is actually picking jobs up. It works with Horizon or a plain queue:work.

Horizon. Each master and its supervisors. A paused or idle supervisor shows up as degraded, a missing master as failing.

The scheduler. Every named scheduled command with its cron expression, last success, last failure and run time. A command that is late, failing, or scheduled but never once finished is caught.

Failed jobs. How many are waiting in the failed-job store, and how old the newest one is.

What lands in the log. Every exception the app reports, the same ones that end up in laravel.log, counted and broken down by exception class and the place in your own code it came from. You hear about them without anyone reading the log. The messages themselves never leave the app.

Outgoing mail. Emails the app could not hand to its mail provider are counted, so a broken mail setup doesn't wait for someone to notice a missing email.

Dependencies. A Composer and npm audit of the lock files actually deployed on the server, re-run whenever they change and at least once a day. Packages that ship to the browser are kept apart from build tools, which never alert.

Release and identity. The commit that is running and when its config was cached, plus a check that the app answering is the one, and the environment, that was expected.

Your own numbers. One line in your code turns anything worth counting into a check, such as webhook failures or bounced emails, with thresholds set in Watchtower like everything else.

Websites. Each address answers with the right status, says what it should, and does it quickly. Plain static sites get this too.

Certificates. Read from the server itself, behind any proxy, flagged two weeks before they lapse, and Urgent three days before.

Provider keys. The email, SMS and AI keys your apps depend on, checked to see they are still accepted. Keys are stored encrypted and never appear in an alert or log.

Account credit. Where a provider reports its balance, a low balance is flagged while there is still time to top up.

A system cannot report its own death. Watchtower checks whether an app is up from outside, over the network, instead of taking the app's word for it. An external heartbeat watches Watchtower.

Silence is never health. A check missing from an app's report counts as unknown, and unknown escalates like any other problem.

Incidents, not messages. Each problem becomes one incident with its own timeline of when it opened, who was told, when it escalated, who acknowledged it and when it resolved.

The monitor decides severity, never the app. Thresholds can be set for the whole fleet, one server or one app, and changing one doesn't require a redeploy.

Servers are watched in their own right. One full disk is one incident, however many apps happen to share it.

Private by design. No health report, incident or alert carries data about a person. Apps report counts, such as how many emails bounced, and nothing about who they were.